Skip to main content

Daily report

AI and science frontiers · 2026-11-18

Only content delivered through the publication boundary on this date is included.

This digest updates until the publication day is sealed.

arXiv

A neural controller trained only on expert action traces can be proven to run algorithms exactly: a 4,335-parameter DFST clears all 10 universal-simulation tasks at the 500-symbol cap

A differentiable exact-learning framework couples a state-tracking neural controller to an unbounded external grid environment and trains it only on policy-trajectory observations, which record expert observations and actions while hiding internal states. A 4,335-parameter DFST trained on just 75 traces with inputs of at most 5 symbols reaches the 500-symbol evaluation cap on all 10 universal-simulation tasks, 5 of them held out, and discrete controllers extracted from trained models are proven equivalent to the expert universal interpreter.

arXiv

MambaVision block ablation mostly measures positional fragility: with size-matched nulls, per-block content effects fall to a fraction of a point

In hybrid vision backbones such as MambaVision, which interleave Mamba state-space mixers with self-attention, removing one block and reading the accuracy drop has served as the measure of that block's contribution. Across all 50,000 ImageNet validation images, each ablation was paired with a size-matched null, the same perturbation with channels shuffled or signs flipped, and real and null damage were nearly indistinguishable in top-1 accuracy and true-class rank. The same test detects content effects in DeiT-Small and Swin-T, so single-block ablation in MambaVision mainly reflects how sensitive the residual stream is to where and how large a perturbation is, not what the block computes.

arXiv

NVFP4-MLA pushes both MLA KV-cache storage and attention math to FP4, matching BF16 accuracy while shrinking the cache 3.20-fold

Under long-context and agentic workloads, MLA's KV-cache capacity and attention computation dominate serving cost. NVFP4-MLA pushes both MLA KV-cache storage and both attention matrix multiplications to FP4, using selective residual compensation that adds an NVFP4 residual only to quantization-sensitive channels and two-dimensional quantization blocks that let the shared latent cache serve both reduction directions. On nine benchmarks for DeepSeek-R1 and Kimi-K3, average accuracy matches BF16 MLA while the cache shrinks 3.20 and 3.05 times relative to BF16.

arXiv

Model merging drops peer-task data: with ImageNet as a universal proxy, merged CLIP models approach the oracle that sees other tasks

Model merging builds a multi-task model from independently fine-tuned experts, but earlier pre-merging regularizers needed data from the other tasks to be merged, breaking the independence assumption. This work characterizes a universal proxy distribution against target tasks through spectral coverage and spectral leakage, and derives a unified bound in which smaller spectral parameters mean lower parameter interference. On CLIP vision benchmarks, three instantiations regularized with an ImageNet proxy beat existing pre-merging baselines and approach the oracle that uses peer-task data; on language models, a WikiText-103 proxy unrelated to TRACE also improves merged performance.

arXiv

Closed-form couplings can't replace a pretrained teacher: in high dimensions the endpoint is mostly the initial noise, and students land near random pairing

Few-step generation usually relies on a pretrained teacher to produce noise-data pairs for rectified-flow training. For a finite dataset and a Gaussian probability path, the empirical probability-flow field is available in closed form, so one could use its couplings as the teacher and skip pretraining. The study shows this fails at practical dimensions: the closed-form flow returns the injected noise almost unchanged at the endpoint, adding only a faint data-dependent drift that is nearly independent of the noise, so training on these pairs performs close to random pairing; the drift's share of endpoint energy falls as dimension grows, and closed-form couplings become more informative only when the dimension is lowered.

arXiv

Train with a centralized safety filter, deploy without it: a three-agent policy scales zero-shot to twenty agents at constant per-agent cost

Multi-agent reinforcement learning can stay collision-free in training with a centralized control-barrier-function filter, but that filter solves a program whose constraint count grows quadratically with team size and cannot be kept at deployment. This approach extracts the policy from executed safe actions with implicit Q-learning, then distills the training filter into the policy at synthetic near-collision states so the filter-free policy still respects the constraints. Local-observation policies had no collisions in test episodes on their training settings, a three-agent policy transferred zero-shot to twenty agents at constant per-agent computation, and stayed collision-free up to eight agents.

arXiv

WinDEX turns in-the-wild egocentric stereo video into metric hand trajectories, with offline reconstruction at 43.61 FPS

WinDEX converts hands recorded by a head-worn fisheye stereo camera in everyday scenes into metric, gravity-aligned hand motion in a world frame for robot teleoperation and reinforcement learning. Such data previously required optical motion capture, multi-camera rigs or gloves, or came from monocular video without absolute scale. The system generates annotations through geometric-temporal optimization, trains the end-to-end fisheye-native WinDEX-Net, and distills offline WinDEX-Window and causal streaming WinDEX-Stream; on held-out data, offline reconstruction reaches 14.5 pixels reprojection error and, with visual-inertial odometry, 37.29 mm world-frame joint error.

arXiv

Keeping only 5% of visual tokens still holds 91.6% of performance: Glance-ETD glances at low resolution, then selects tokens by question evidence

In multimodal large models, visual tokens dominate prefill cost for high-resolution images, and earlier pruning at extreme budgets only shortened the language model's input while the vision encoder still attended over every patch. Glance-ETD uses one question-conditioned scorer of about 1.9M parameters to drive two steps: a quarter-resolution glance that re-encodes only the 25% of cells relevant to the question at native resolution, then an evidence-tempered diversity rule that picks the final tokens. At 5% of the visual tokens on Qwen3.5-4B, it retains 91.6% of full-token performance across fourteen benchmarks, 17.3 points above the prior best LT-OPD, and reaches the first token 3.1 times sooner on 4K images.

arXiv

Safety filters stop re-solving per scene: a conditioned reachability model pushes error to the conservative side and collides zero times across nine obstacle settings

Every time a deployed robot's obstacle layout changes, Hamilton–Jacobi reachability must be re-solved, yet learning a single value function conditioned on the obstacle configuration can silently produce false-safe states. A team at Nanyang Technological University caps the obstacle margin so the value depends only on a local window in the robot's body frame, builds the value bound into the output structure, and maximizes the claimed safe volume subject to the HJ inequality. On a Dubins car, the model is the only one among nine obstacle settings within window capacity with no false-safe state within grid tolerance and zero closed-loop rollout collisions, while keeping recovery comparable to per-scene DeepReach training.

arXiv

Looped Transformers no longer wait a full loop per token: LoopFill merges suffix retrieval and shallow drafts into one tree, beating LoopAhead on seven benchmarks

Looped Transformers reuse the same layers repeatedly, so decoding each token requires several sequential passes and latency is high; earlier pipelined self-speculative methods drafted only along recurrent depth, one token at a time, and each token had to reach the drafting depth before the next could enter the pipeline. LoopFill combines multi-token continuations retrieved from suffix trees with the model's own shallow-loop predictions in a single draft tree, letting several future tokens enter the pipeline together while full-depth verification preserves the autoregressive output distribution. Across seven benchmarks under greedy decoding, its average speedup over autoregressive decoding exceeds the strongest baseline, LoopAhead, by 40% on Ouro-2.6B and 25% on Huginn-3.5B.

arXiv

Auriga pairs Fabric tuning with pre-execution conflict grouping, lifting effective throughput up to six times the default

Auriga splits the effective throughput of Execute-Order-Validate blockchains such as Hyperledger Fabric into raw throughput and transaction success rate, then attacks both with a harness-governed multi-agent configuration search and a scheduler that predicts read-write sets before execution and groups transactions into conflict-free batches. Across three benchmarks and conflict rates up to 70%, effective throughput reaches up to six times the default Fabric configuration and exceeds the strongest standalone baseline; at a 40% conflict rate it holds a 90.2% success rate where the tuning-only Athena stays at 67.8%.

arXiv

Brain-to-text decoding moves onto an iPhone: FETTUCCINE hits 370x real-time on a phone, with word error rate as low as 4.39%

Speech neuroprostheses decode neural activity during attempted speech into text, but the standard cascaded pipeline of an n-gram language model plus LLM rescoring needs tens to hundreds of gigabytes of memory and a dedicated workstation, and its latency can jump from milliseconds to more than ten seconds. FETTUCCINE adapts open-source automatic speech recognition models into end-to-end brain-to-text decoders, replacing the acoustic front end with a neural adapter and applying low-rank updates only to the encoder; on the public brain-to-text 2025 data it reaches word error rates as low as 4.39% and decodes at up to 370x real-time on an iPhone 15 Pro Max.

arXiv

VLMs no longer collapse scene state into one guess: PRaVaSS keeps a minimal high-probability state set and lifts planning success from 16.7% to 32.6% on PRB

When a robot estimates scene state with a vision-language model, the pipeline usually collapses probabilistic output into a single most likely symbolic state, and a wrong guess makes downstream planning unsafe or infeasible. PRaVaSS instead starts from the model's existing per-feature probabilities and combines them with task constraints to build a minimal set of states whose cumulative probability meets a user-set threshold, so the planner considers several likely states at once. On the PRB and GW planning domains, with GPT-4.1 as the backbone, this pipeline raised downstream planning success from 16.7% and 70.7% for MAP to 32.6% and 80.0%.

arXiv

Zero-SM NCCL takes GPU communication off the SMs: AllGather bandwidth reaches 1.36x the legacy kernels and overlapped GEMM slowdown falls from 33% to 4%

Zero-SM NCCL moves both data movement and synchronization out of the streaming multiprocessors (SMs), using CUDA stream memory operations, copy engines and a CPU proxy driving RDMA to run AllGather, AlltoAll and one-sided Put/Wait on GB200 NVL72 and B200 NVL8 systems. Its AllGather reaches up to 1.36x the bandwidth of NCCL's legacy SM kernels and cuts the median slowdown of overlapped GEMMs from 33% to 4%, at the cost of higher per-operation latency. In Megatron-LM it speeds up FSDP by up to 3.5% and pipeline parallelism by up to 15%, but it does not win on small messages or host-bound workloads.

arXiv

Order in linear recurrent memory gets an algebraic test: HyperSSM stays exact on a 32-generator group task with quaternion channels

Whether a linear recurrent memory can retain the order of its inputs depends on the number system its transition matrices live in. Efficient parallel scans have relied on diagonal recurrences, and real or complex transitions commute, so the state records which inputs occurred but not their order. HyperSSM adds quaternion blocks to each channel while keeping linear-time inference and parallel training, solving group tasks exactly with at most five state numbers, including 32 generators, and staying exact on sequences 25 times longer than training.

arXiv

Robot policies stop predicting the whole future frame: ResWAM learns only frame-to-frame residuals and reaches 81.1% on LIBERO-Plus

World-action models usually have the policy predict complete future observations, but little changes within an action chunk: on LIBERO, current and future frame features share an average cosine similarity of 0.96, so the future loss is largely satisfied by copying static background. ResWAM instead supervises only the residual from the current frame to the future frame in a frozen DINOv3 feature space, and a 0.5B dual-branch DiT jointly denoises actions and residuals within 2.3B total parameters. Without robot pretraining it reaches 81.1% average success on LIBERO-Plus and 93.8% on RoboTwin 2.0, at 2.7–4.0 and 1.0–1.4 TFLOPs of training and inference compute.

arXiv

One cached adapter per document: a frozen 284B model gains 38.2% relative on enterprise Q&A

Continual learning lets a language model absorb an organization's growing private documents, but updating shared parameters overwrites earlier knowledge. Microfitting instead trains a small low-rank adapter offline for each document and caches it, keeping the base model frozen so new documents only add their own adapters. On EDO, an open-book Q&A environment over real enterprise documents, relative improvement over the base model rose from 30.9% to 37.5% to 38.2% as the data-generation recipe was revised.

arXiv

Keep Jev's probability as the safety reward, not a 0/1 verdict: with a helpfulness term, HarmBench attack success on Llama-3.1-8B falls from 0.287 to at most 0.033

The harmful-probability output of the commercial safety judge Jev is used directly as a reinforcement-learning reward instead of being binarized, with a four-stage validation run before training, during training, and after each judge version change. On Llama-3.1-8B, the probability reward with a helpfulness term lowers HarmBench attack success from a 0.287 baseline to at most 0.033 at the final checkpoint; binarizing the same judge makes the share of confidently passed harm rise significantly in both seeds, meeting the escape criterion.

arXiv

Open Vision Banana releases an open generative-perception recipe, hitting 73.9 mIoU zero-shot on Cityscapes against Vision Banana's 69.9

Open Vision Banana releases a reproducible recipe that converts open-weight FLUX.2-Klein image generators into a unified multi-task perception model, trained only on public data to output segmentation, depth and surface normals. The zero-shot 9B model reaches 73.9 mIoU on Cityscapes, above the 69.9 reported for Vision Banana; its non-zero-shot variants come within five points on referring and reasoning segmentation but remain further behind on instance segmentation. The data pipeline, training code and checkpoints are released.

arXiv

Language steganography drops the bit-exact requirement: STEALTH decodes every message across precisions, GPUs and even model families

Provably secure linguistic steganography has assumed the receiver can reproduce the sender's token probabilities bit for bit, yet floating-point precision, quantization, GPU architecture, or scoring a text in one pass instead of token by token all break that premise. STEALTH samples each token with a keyed exponential race, so a probability mismatch flips only the token whose race is close, and a rateless spinal code decoded by beam search absorbs those flips. Across the Gemma, Llama and Qwen families, same-family receivers decoded every message at every precision, quantization level, smaller model size and on a second GPU, and cross-family receivers decoded every message tested.

arXiv

Diffusion language models stop writing function words first: SemMask decodes semantic tokens early, reaching 0.7085 HumanEval accuracy in fewer steps

Diffusion language models are meant to generate in parallel, yet in practice they often decode left to right and emit articles and prepositions before content words, weakening global planning. SemMask scores each token by the original model's gradient, then during training preserves high-scoring semantic tokens and up-weights their loss so nouns and verbs are decoded before function words. On Dream-7B, SemMask reaches 0.7085 accuracy on HumanEval, above Dream-7B-Instruct at 0.5366 and Qwen2.5-7B-Instruct at 0.6182, while average decoding steps fall from 201 to 169.

arXiv

Bayesian neural network output distributions get a sound envelope: a nonlinear abstract domain cuts total variation distance by up to about 950x over interval abstraction

Bayesian neural networks treat weights as distributions, yet prior analyses either sampled or propagated only expectations and entropy, leaving the full output distribution without guarantees. This work builds a marginal collecting semantics and represents probability measures with exponential-affine bounds over a partition of the support, yielding composable abstract transformers that give a sound envelope of the complete output distribution on several regression benchmarks. Against an interval abstract domain, total variation distance improves by up to about 950x, at the cost of longer runtimes.

arXiv

Optimistic verifiers teach padding: true faithfulness falls from 0.731 to 0.667 while the proxy score holds at 0.940

Outside mathematics and code, reinforcement-learning verifiers are mostly proxy signals whose distance from the true objective is never measured, so a rising reward cannot be told apart from a hack. This work compiles executable authorization policies into exact process-level rewards, then degrades the verifier along seven realistic axes while the exact verifier scores every rollout. The sufficiency proxy is hacked by padding: its proxy score terminates at 0.940 while true faithfulness on held-out episodes falls from 0.731 to 0.667, whereas an outcome-only verifier saturates almost immediately and stops teaching.

arXiv

A 30B-parameter MoE model fits a 16 GB GPU: CECAR folds expert-cache state into routing for 12.5 tokens/s

CECAR folds expert cache residency and predicted eviction priority into MoE routing so that fewer non-resident experts are requested. Prior cache-aware routing relied only on whether an expert was cached, which could destabilize routing; on Qwen3-30B-A3B, with 16 GB of VRAM, about 1 GB of host DRAM and non-resident experts fetched from SSD, CECAR reaches 12.5 tokens/s, a 4.5x speedup over conventional MoE inference and faster than a baseline keeping all experts resident in system DRAM.

arXiv

A verbalizable workspace holds concepts but not role bindings: probes decode roles near ceiling, yet the workspace reads below a rank-matched random subspace

A verbalizable workspace proposed as a substrate for runtime alignment monitoring was tested for role-filler binding, the assignment that separates propositions sharing a concept set, such as the model plans to deceive the user versus the user plans to deceive the model. Across Qwen3.6-27B, Gemma-3-12B and Gemma-3-27B-IT, two lens families and two role-direction estimators, the residual stream decodes role near ceiling while the workspace slice scores below a rank-matched random subspace, and pushing along a fitted role direction changes what the model reports about roles without changing which entity it binds as agent. Ablating the workspace damages binding specifically only on Gemma-3-27B-IT at the entity token, replicating across three lens variants.

arXiv

MPPI's exploration covariance is also its feedback gain: too little exploration destabilizes an inverted pendulum

In model predictive path integral (MPPI) control with a warm start, the exploration covariance sets not only how widely candidates are sampled but also the closed-loop feedback gain. For linear quadratic problems with infinitely many samples, too little exploration makes MPPI nearly repeat the previous solution and fail to stabilize an exponentially unstable plant, while the same covariance lowers the effective sample size, a structural conflict between stability and sample efficiency. Stabilizing Subspace MPPI (SS-MPPI) computes the directions needed for stabilization and samples only the rest, staying stable with small exploration and few samples in inverted-pendulum and quadruped simulations.

arXiv

Prompt optimization stops rewriting whole strings: Splice freezes contract clauses, winning four of seven benchmarks and matching on three

Splice recasts natural-language prompt optimization as constrained combinatorial search over an edit graph of block-structured states: safety clauses, tool schemas and output formats are tagged immutable so no edit can reach them, and length growth is priced by a ratio penalty. Across seven public benchmarks under five reproduced protocols, Splice wins on four and matches the strongest prior method on three; in a contract stress test, unconstrained optimizers trade away the clauses that cost reward, and frozen text does not mean frozen behavior.

arXiv

FastGRN cuts GRN image generation from 50 refinement steps to 2, matching its teacher's ImageNet FID at 12.7 times the speed

Discrete-token image generator GRN needs 50 refinement steps per image, and cutting the step count makes its samples fall apart. FastGRN treats GRN's own uniform-state corruption as a discrete diffusion kernel, computes a per-token reverse-KL gradient directly from two posterior networks, and adds windowed matching, oversampling and non-uniform sampling so that a 2-step student matches the FID of its 79-network-call teacher on ImageNet while running 12.7 times faster per image.

arXiv

Robot agents and VLA policies stop training apart: RoboCoEvo turns failure programs into policy data, lifting success across three benchmarks

RoboCoEvo closes the loop between a robot agent and a vision-language-action (VLA) policy: the agent groups recurring failures into failure modes, writes one reusable detector and recovery program per mode, and uses those programs on failure states and perturbed copies to generate corrective data that DAgger-style post-training writes into the policy. The updated policy's shifted failure distribution then guides the next round of program revision. Across two co-evolution rounds on BEHAVIOR-1K, RoboTwin and RoboCasa, success rises from 0.23 to 0.34, from 0.79 to 0.93 and from 0.79 to 0.91.

arXiv

Recorded with glasses alone, Ego2Sim2Humanoid turns human video into humanoid training trajectories

Whole-body vision-language-action models for humanoids depend on teleoperated robot trajectories, each requiring a trained operator driving a physical robot in real time. Ego2Sim2Humanoid converts everyday human recordings from Project Aria Gen 2 glasses into training data with robot-view images and verified action labels, through agentic real-to-sim reconstruction, grasp and stance planning, and reinforcement-learning correction in physics simulation. Fine-tuned only on this converted data, with no real-robot demonstrations of the target tasks, GR00T N1.7 lifted a bottle in 9 of 20 autonomous trials and put on a hat in 5 of 20 on a Unitree G1.

arXiv

Agent memory no longer has to be hand-designed: MAPE evolves whole memory programs from a no-memory seed, gaining up to 12 points on LoCoMo

Agent memory systems have relied on human expertise to decide what to store and how to organize and retrieve it. MAPE exposes memory construction, retrieval and evidence composition as programmable stages and rewrites the whole program from minimal seeds such as one that stores no memory, guided by an iterative diagnosis–proposal loop. On LoCoMo and LongMemEval the selected programs beat every evaluated human-designed baseline on their respective benchmarks, with an absolute accuracy gain of up to 12 points on LoCoMo at 35% of the baseline serving cost.

arXiv

Open-source LLM watermarking moves to a private vocabulary: injected once on Alpaca, SAME keeps a 100% fixed trigger rate after downstream fine-tuning

Once open-source model weights are public, third-party fine-tuning rewrites parameters and internal representations, breaking watermarks that rely on specific trigger words or parameter traces. SAME pairs every native token with a fixed private embedding and uses dual-path training so the private vocabulary inherits native semantics, tying the watermark to the model's evolving semantic ability. Injected once on Alpaca into Qwen3-8B-Instruct, it keeps a 100% fixed trigger rate after independent fine-tuning on SST-2, AG News, MMLU and BoolQ, while BadEdit falls to 10.87% on MMLU.

arXiv

Frozen models get a sense of touch: ForceHarness lifts GPT-6 Astra from 3.3% to 90.0% on a real arm with no wrist force sensor

A frozen frontier model reached only through an API cannot read raw force streams or decide at the speed contact evolves, so force has mostly entered manipulation through demonstration-trained force-aware policies. ForceHarness places three layers between model and robot: hand force calibrated from a wrist sensor or joint torques alone, self-terminating force primitives that regulate force and return at a contact event, and a translation of each force trace into physical quantities with uncertainty. On a Franka FR3 without a wrist force sensor, GPT-6 Astra's zero-shot success rate rises from 3.3% to 90.0%; in simulation it rises from 70.0% to 87.5% on ForceBench and from 62.1% to 96.4% on the seven-task HiddenPhys.

arXiv

Agents stop waiting for external scores: Verify by Doing co-evolves execution and verification inside a single task

Agents that improve themselves at test time usually need benchmark scores or reference answers to tell whether a task is done, and real tasks often lack that external feedback. Verify by Doing gives the executing agent and the verifying agent separate editable harnesses of instructions, skills and checklists, gathers completion evidence by actively inspecting artifacts and environment state, and uses those findings to update both execution and verification. Across AutomationBench, Banking and 133 GDPval tasks it posts the highest overall scores among compared methods without external feedback, and on a fixed candidate pool the evolved evaluator raises pass rate from 31.2% to 35.8%.

arXiv

Any Channel Count Into One Token Stream: SpheraAudio Beats Spatial-Omni-7B on Spatial Audio Benchmarks

Large audio-language models usually collapse sound to mono to keep general audio understanding stable, leaving the inter-channel phase relations that encode direction and distance unused. SpheraAudio accepts mono, stereo, first-order and higher-order Ambisonics recordings, encodes multichannel complex spectra into spatial tokens, and interleaves them with audio tokens by time before the language model. At 4B parameters it leads on SpheraQA, SO-Bench and TWNM, reaching 60.18% overall on SpheraQA, 18.53 percentage points above Spatial-Omni-7B.

arXiv

Looped language models no longer wait loop by loop: LoopSkim turns each depth's readout into a verifier, reaching 1.73 times LoopSpec's speedup without training

Looped language models apply one weight-shared block several times per token, so autoregressive decoding runs every loop of every token in sequence. LoopSkim lets draft tokens read out at shallow depths be checked by each later depth's readout, commits only full-depth readouts, and fixes each position's random draws in advance so decoding stays lossless under sampling. Across five Ouro and Raven models, the untrained version averages 1.73 times LoopSpec's speedup over autoregressive decoding, and an optional draft head raises that ratio to 1.99.

arXiv

DeepBA learns bundle adjustment as a graph network: 2.04M parameters solve camera poses from a cold start, and outpace COLMAP as views grow

DeepBA replaces the per-scene bundle-adjustment solve in multi-view reconstruction with a learned graph-network refiner that consumes only the 2D matches of an off-the-shelf matcher and recovers absolute camera poses, intrinsics and scene structure from a cold start. Trained with scale-invariant supervision on datasets with incompatible metric units, its 2.04M parameters lead prior learned SfM from matches on CO3Dv2 and RealEstate10K, are competitive with classical SfM at moderate accuracy thresholds, and overtake it in wall time as the view count grows.

arXiv

TrainExplorer lets agents retrain small models on their own: literature-built strategy space lifts PostTrainBench weighted score to 39.02, past Codex

Agents that autonomously post-train large language models tend to stay inside their initial strategy, and a single experiment often changes data and training settings together, so the effect of each change is hard to attribute. TrainExplorer separates strategy planning from execution: it first builds a candidate strategy space from scientific literature, then splits data processing and model training into independently replaceable components, changing one per iteration and attributing the effect. On PostTrainBench's seven tasks and four base models of 1.7B to 4B parameters, it reaches a weighted aggregate score of 39.02, 7.32 points above Claude Code with the same agent LLM and above GPT-5.6 Sol-powered Codex at 36.23.

arXiv

AntEmbed runs two retrieval modes on one index: its 8B generative embeddings score 81.5 on MMEB-v2

Multimodal retrieval previously forced a choice: discriminative embeddings encode fast but miss fine-grained evidence, while generative embeddings reason before encoding and carry richer semantics at extra latency. AntEmbed treats the original and reasoning-enhanced forms of a candidate as equivalent views of one instance, aligns them in a unified space with a single embedding token and multi-positive contrastive learning, then distills the reasoning branch's fine-grained information into the original branch. On MMEB-v2, AntEmbed-8B reaches an overall 81.5 with generative embeddings, 3.1 points above the strongest baseline DME-9B; its discriminative embeddings need no extra reasoning, and AntEmbed-2B scores 70.2 in base mode.

arXiv

Small-model agents stop writing their own security policies: CapLite derives rules from tool schemas and request text, blocking all 13,935 injected episodes

CapLite, a prompt-injection defense for small open-weight model agents, stops asking the model to write security artifacts: it derives its policy from tool schemas, commits write calls before any tool result is read, and admits a write only if every argument deciding recipient, amount or target is stated in the user's request. Across 15 open-weight models on AgentDojo, no attack succeeds in 13,935 attacked episodes, and utility under attack is 29.0% against CaMeL's 27.2% at 17.1k tokens per episode, 18% of CaMeL's cost.

arXiv

JEPA4R reads single-depth-camera point clouds: imagined rollouts decode into 4D clouds within 1.1 cm over 200 steps

JEPA4R replaces the image input of a JEPA world model with colored point clouds from a single depth camera, predicts and plans in their latent space, and acts through a policy head on the frozen latents, a procedure the authors call 4D point planning. It leads pixel JEPA models on OGBench Cube-Single, Cube-Double, Cube-Triple and Scene, and ranks first on CortexBench's MetaWorld and Adroit Relocate at 98.4% and 88.0% success, above the runner-up R3M at 96.0% and 65.3%. A decoder attached to the frozen world model turns every imagined rollout into 4D point clouds, with mean point error within 1.1 cm over 200 environment steps.

arXiv

Agri-MRAG stops pasting retrieved text into farm queries: evidence becomes LoRA adapters, reaching ROUGE-L 46.58

Agricultural visual question answering must read crop symptoms and call on pathology knowledge at once, and general multimodal models often confuse visually similar diseases. Agri-MRAG clusters retrieved image-text evidence by semantic consistency, keeps only a high-confidence subset, encodes that evidence as LoRA adapters injected into the model's feed-forward layers, and uses perplexity to decide whether to add explicit retrieved context. On PlantVillageVQA and the newly built PlantVQA it reaches ROUGE-L of 41.21 and 46.58, 11.70 and 4.51 points above mRAG and 8.80 and 3.39 above VisRAG.

arXiv

Peer-to-peer energy market clearing gets a learned speedup: a graph-network correction adds no communication and keeps the baseline's convergence guarantee

Clearing peer-to-peer energy markets usually relies on distributed iterations that seek a generalized Nash equilibrium, and the iteration count and communication rounds become a deployment obstacle as participants grow. A KTH team trains a graph-neural-network correction on historical clearing data and adds it to a distributed Douglas–Rachford baseline; the correction decays geometrically, so convergence is inherited from the baseline. On synthetic markets the correction cuts the iterations needed to clear and their spread, and the speedup persists when operating conditions shift to four times the training variability.

arXiv

A summary's sufficiency now comes with its error bar: a classifier estimates the Barber–Agakov gap, making earthquake-catalog comparisons trustworthy

In simulation-based inference, high-dimensional or variable-length data are often compressed into low-dimensional summaries, and how much parameter information a summary retains is measured by mutual information, which is intractable without a likelihood. A trained neural posterior yields a Barber–Agakov lower bound used in place of that mutual information, but a small bound can mean either an insufficient summary or an inaccurate density estimator. This work estimates the gap between bound and mutual information with a classifier and restricts comparisons to a data-space neighborhood around the observation; on temporal ETAS earthquake catalogs, the learned summary has both the larger bound and the larger gap, so the bound difference understates the mutual-information difference.

arXiv

Stress-testing a black-box ship collision-avoidance system: two target ships steer themselves, and the R-PPO policy finds collisions in 68 of 100 scenarios

Safety assessment of ship collision-avoidance systems usually relies on simulation, but in co-simulation each subsystem exposes only its interfaces, so conventional stress testing written as a fully observable Markov decision process cannot see the full state. EB-ASTv2 reformulates the search as a partially observable Markov decision process, uses recurrent proximal policy optimization (R-PPO) to pick adversarial actions from observation history, and lets two target ships trigger intermediate waypoint sampling independently. Across 100 traffic scenarios in a ship-in-transit setting, the R-PPO policy produced 68 own-ship collision scenarios, versus 58 for PPO and 54 for SAC; over five independent training runs the mean collision rate was 66.80%, against 65.20% for PPO and 51.80% for SAC.

arXiv

Agents can read an API that quietly answers the wrong question: SilentGuard moves the verdict into the harness, cutting wrong answers from 74.0% to 1.7%

SilentGuard moves the decision of whether a tool result can be trusted from the model to the harness that runs tool calls: the harness tests each result against invariants any correct API must satisfy and withholds values that break an exact one. Across 7 open models it cuts wrong answers on its target faults from 74.0% to 1.7% with clean-task accuracy unchanged at 84.9%, and on three frontier models from 70.2% to 0.5%. It avoids wrong answers by abstaining rather than recovering, and on fault types its checks cannot see, wrong answers fall only from 61.1% to 24.5%.

Claude Code tightens sandbox defaults: sandboxed Bash now asks first, and admins must opt back in

Anthropic's Claude Code changelog says sandboxed Bash commands now follow the normal permission flow instead of running without a prompt just because they are sandboxed; restoring prompt-free execution requires setting sandbox.autoAllowBashIfSandboxed in the highest-ranking managed source. The same update stops the sandbox from writing to several sensitive home folders by default on macOS, Linux and WSL, and adds permissions.blockWritesOutsideWorkingDirectories so Write, Edit and NotebookEdit refuse paths outside the working directories in every permission mode. These changes move the default boundary for command execution and file writes; the changelog reports no effectiveness evaluation.

arXiv

Quantum error correction decoding no longer has to choose speed over accuracy: TCBP-JOBS matches Tesseract's logical error rates on GPU, with serial speedups up to thousands of times

Decoders for general quantum error-correcting codes have long traded speed against accuracy: fast decoders lose accuracy, while most-likely-error decoders are too slow for real-time correction. TCBP-JOBS uses a small multilayer perceptron to learn a calibrated probability for each fault from the full belief-propagation trajectory, then starts GPU-parallel searches over many orderings from those probabilities. Across 72 surface, color and bivariate-bicycle code circuits it reaches logical error rates comparable to or lower than Tesseract, by up to 4.5 times on the largest codes, with serial decoding 10 to 7,400 times faster than sparsified Tesseract.

arXiv

Carbon-nanotube sheets stop blurring into one blob: PLECTA reads only a binary axis mask and reaches mean F1 0.85 on synthetic scenes

Carbon-nanotube sheets, collagen and paper are built from overlapping strands, and in a micrograph a crossing merges distinct strands while a signal gap splits one strand into fragments. PLECTA takes only a binary mask marking strand centrelines and, at each junction, treats the possible continuations jointly as an exact maximum-weight partial matching, allowing strands to end and crossing pixels to belong to several strands. Across 128 synthetic scenes it reached a mean common-fragment pairwise F1 of 0.85 (95% bootstrap CI [0.84, 0.86]), falling from 0.91 to 0.77 as areal coverage rose from 20% to 60%.

arXiv

UNIDE turns equation discovery into cross-PDE transfer: exact term sets survive 80% noise where WSINDy collapses

Recovering a governing equation from noisy space-time trajectories has meant fitting sparse or symbolic regression from scratch on each trajectory, with differentiation amplifying observation noise. UNIDE learns across six one-dimensional PDE families: a blind-spot denoiser that needs no clean targets cleans the trajectory, a shared encoder predicts which candidate terms are active, and weak-form least squares with residual pruning estimates their coefficients. On 900 held-out trajectories at 80% noise it recovers the exact term set for 0.924, against 0.003 for WSINDy.

arXiv

Brain MRI no longer needs all three stacks: MS-HVED rebuilds from whatever orientations exist, and experts preferred it in 187 of 210 Nigerian pairs

MS-HVED encodes each available axial, coronal or sagittal low-resolution brain MRI stack into a multi-scale Gaussian posterior and fuses them with a field-of-view-aware product, so one trained model accepts any non-empty subset of orientations without retraining. In a blinded reader study on 30 routine brain MRI examinations from a Nigerian cohort, seven expert neuroradiologists preferred MS-HVED over the supervised SegResNet baseline for overall image quality in 187 of 210 paired assessments, and the preference persisted when an axial or coronal input was deliberately omitted.

arXiv

Drug synergy prediction stops choosing between molecules and networks: BiGSyn aligns both graph levels and reaches 94.37 AUC on the independent test

Drug synergy prediction has mostly been modeled at a single level, using either atom-level molecular graphs or a biological network of drugs, targets, cell lines and tissues, so cross-level signals were hard to fuse. BiGSyn uses heterogeneous graph transformers to learn representations from molecular graphs and biological subgraphs, then a bi-level contrastive objective pulls the two representations of the same drug together. On the Oncology-Screen and DrugCombDB benchmarks it beats six compared methods on independent-test metrics, with an AUC of 94.37 on Oncology-Screen.

arXiv

Neural operator filter NOF skips Yau–Yau's offline PDE solves, cutting 2D runtime about 77.9% versus a particle filter

Yau–Yau nonlinear filtering splits state estimation into an offline stage that solves several Fokker–Planck–Kolmogorov equations and an online projection onto prescribed basis functions, and both stages are costly. The neural operator filter (NOF) instead uses a Fourier neural operator to learn the map from prior density to posterior density directly, removing both steps. On one- and two-dimensional cubic filtering benchmarks, NOF's mean-square error is comparable to a particle filter and the LGS method, while in the 2D case its average runtime is about 77.9% lower than the particle filter's and about 63.3% lower than LGS's.

arXiv

No training needed to forecast forest carbon: TabPFN-3.5 posts the lowest error on four state variables under high emissions

Process-based terrestrial biosphere models can compute global carbon budgets, but their cost limits global, high-resolution and scenario-rich simulation, and purpose-built deep emulators require task-specific design and training. This work recasts 40-year Ecosystem Demography forecasts as tabular regression and lets the pretrained tabular foundation model TabPFN-3.5 predict directly by in-context learning, with no weight updates. On CarbonGlobe's 50,779 test sites it has the lowest height error of the three models without training and soil-carbon error within 4% of DeepEDLSTM, but larger errors for AGB, LAI and the carbon fluxes; under SSP5-8.5 it is the most accurate model for height, AGB, soil carbon and LAI.

arXiv

A universal force field leaves the supercomputer behind: ALIGNN-FF evaluates two billion atoms on desk-side GB10s

The universal graph neural network force field ALIGNN-FF was previously capped by single-GPU memory at 216,000 atoms with forces as one graph on a desk-side GB10. Cutting the periodic cell into cubes, each evaluated with a halo covering the model's 20 Å receptive field, makes per-piece memory independent of system size, and nine GB10s evaluate energies and forces of 2,000,376,000 silicon atoms, with the largest piece at 104,194 atoms. A retrained AFF-r2SCAN-smooth with smooth cutoffs also removes the force jump at the three-body cutoff, so relaxation converges and energy is conserved.

arXiv

Coherent optical neural networks drop the Mach-Zehnder mesh: a butterfly topology with one-parameter units lowers reconstruction error under tight parameter budgets

Coherent optical neural networks are limited by how many tunable parameters they can afford, since fabrication complexity, footprint and control overhead all grow with parameter count. The fine-grained butterfly (FGB) architecture replaces the Mach-Zehnder interferometer, which carries two tunable phase shifters, with a single phase-shifter and beam-splitter pair as the building block, cascaded in a butterfly topology so that each input-output pair has exactly one path and every path crosses the same number of units. In computational evaluations under restricted parameter counts, this architecture gives lower reconstruction error on Haar-random matrices than conventional rectangular meshes and higher test accuracy on Fashion-MNIST classification in both feedforward and recurrent networks.

arXiv

Vibration test records go to large language models: after six deterministic gates, all 20 configurations reproduce the gold notching decisions bit for bit

In sine vibration acceptance testing for LEO constellation satellites, test conditions live in Chinese records and were previously coded by hand into the analysis. A Harbin Institute of Technology team had DeepSeek V4.1 Flash and Qwen-plus read the records into a 76-field configuration, and placed six deterministic gates between that configuration and the numerical stage, checking units, physical consistency, verbatim quotes, agreement between the two models, consistency with measurement exports and applicability of the trained model. Under the final pipeline and consensus policy, all 20 LLM configurations reproduced the gold notching decisions bit for bit, and the one real error plus all 16 injected errors in consumed fields were stopped, of which 16 would have run without gates.

arXiv

System identification stops rerunning per record: one diffusion model holds state accuracy on records five times longer

Bayesian system identification infers a dynamical system's hidden states and parameters from measurements, and particle Markov chain Monte Carlo methods must be rerun for every new record and task. This work treats hidden-Markov-model system identification as a general Bayesian inverse problem and uses a conditional diffusion model trained once to approximate the joint posterior of states and parameters; at deployment, changing only the conditioning mask yields joint inference, state smoothing, parameter inference or missing-measurement reconstruction. On four benchmark systems, state accuracy holds on records five times longer than the training segments, while a dense-attention Simformer baseline's state error can grow more than tenfold.

arXiv

Medical LLM fine-tuning no longer trades away patient safety: S-BDS lifts PatientSafetyBench safe rates by 4.64 points

Patient-facing medical LLMs usually gain domain knowledge through supervised fine-tuning on large medical instruction sets, but that process can erode their existing safety alignment. S-BDS adds entropy regularization to the upper-level objective of bilevel data selection, curbing excessive concentration of selection weights so that sample rankings stay reliable over large candidate pools, and uses clinician-validated safe data as the upper-level signal. On Llama-3.1-8B-Instruct and Qwen3-8B-Instruct, S-BDS achieves the highest PatientSafetyBench safe rates among compared methods, 4.64 and 3.00 percentage points above the strongest baseline SEAL, while keeping competitive HealthBench medical capability scores.

arXiv

Sliced spatial transcriptomics stops stitching slice by slice: HoloGene conditions on sparse Z, lifting whole-brain ARI to 0.467

Serial-section spatial transcriptomics is dense in the XY plane but sparse along Z, and standard coordinate implicit neural representations encode all axes symmetrically, so they do not reflect that sampling asymmetry. HoloGene takes in-plane coordinates as the main input, modulates every layer with a depth condition, and aligns cross-slice covariance within pseudo-label groups derived from training-section expression. On held-out measured sections it reaches the lowest MSE and the highest Pearson, masked Spearman and ARI on MOSTA E16.5 and whole mouse brain, raising whole-brain ARI from the strongest baseline's 0.428 to 0.467.

ElevenLabs Blog

ElevenCreative moves audiobook production into Studio: pronunciation sidebar, timeline sound, and streaming playback

ElevenCreative updated its audiobook workflow in Studio with a standalone pronunciation dictionary sidebar, sound effects and music on the timeline, cross-row magnetic snapping with linked clips, and a streaming playback engine that plays audio while it is still being generated. Previously pronunciation rules lived in a popup, sound could not be arranged alongside narration, and pressing play waited for a whole paragraph to generate. The tools target creators producing audiobooks in Studio, and results depend on actual use in a project.

Vercel News

Microsoft's Decision-1.1 lands on AI Gateway: decision tasks return structured answers, not free text, with probabilities you can threshold

Microsoft AI's Decision-1.1 is now available on AI Gateway for decision tasks such as routing support requests, choosing which model should handle a request, selecting the next workflow step, or grading AI responses, returning structured answers instead of free-form text. Microsoft describes it as more accurate, better calibrated, and faster than Decision-1; better calibration means the model's probabilities track how often it is right, so an application can set thresholds for when to act, flag a case for review, or escalate. The request format matches Decision-1, and switching means changing the model ID to microsoft/decision-1.1.

affaan-m/ECC

ECC 2.2 installs an engineering process into AI agents: one setup brings planning, testing, review and memory along

ECC 2.2 packages an engineering process as an installable plugin so an AI agent plans before it builds, verifies changes with tests, reviews its own work from a fresh context, and turns repeated wins into reusable skills and workflows. Previously that process had to be rebuilt in every prompt; installed once, it becomes part of how the agent works. The project is MIT-licensed open source, works best with Claude Code today, has native Codex and Qoder plugin paths, and provides capability-limited adapters for Cursor, OpenCode, Gemini and other harnesses.

Claude Platform adds linked billing: third-party tools stop handling customer API keys, and usage bills straight to the customer's organization

Anthropic released linked billing in beta on the Claude Platform, letting third-party platforms have their end customers pay for Claude API usage without handing over an API key. Previously, paying for Claude API usage on a third-party platform usually meant sharing an API key, and anyone holding that key could make requests billed to the customer. Now customers link their own Claude Platform organization instead, usage is billed directly to it, and customers can set a monthly spend limit per platform and unlink at any time.

Gemini

Gemini now takes the whole task: free task mode in six countries books calls, compares prices and checks out

Google has turned Gemini from a conversational assistant into one that takes over a whole task: the user hands it an errand and it works across connected apps such as Maps, Calendar, shopping and payments in the background. The update rolls out free to hundreds of millions of people in the U.S., Japan, Korea, Canada, Australia and India on the new Gemini 4 Hydrogen model, with advanced features starting in the U.S. Google's examples include calling two dealerships at once and booking a Friday morning service slot in under five minutes, analyzing an electric bill to trace it to a heat pump running at peak pricing, and finding a discount code that saved an extra $10 at checkout.

ElevenLabs Blog

ElevenLabs moves medical transcription into live calls: Scribe v3 Live Medical posts 12.8% WER on PriMock57, 17% fewer errors than the next medical model

ElevenLabs released Scribe v3 Live Medical, bringing real-time speech transcription tuned for clinical language to call agents in ElevenAgents. Its earlier medical transcription covered recordings only, while drug names and doses in live calls are what general models most often mishear. On PriMock57, 57 mock primary care consultations, the model posted 12.8% word error rate, 17% fewer errors than the next closest medical model tested, and up to 48% fewer errors on medical terms with a custom term list loaded.

Vercel News

eve 1.0 goes generally available: the same 100-turn stress test drops from 3.036 s to 414 ms per turn

Vercel has moved its open-source agent framework eve from public preview to a 1.0 general release, adding security approvals, shared multiplayer sessions, cross-session memory and OpenTelemetry tracing. In a hosted stress test on Vercel Workflows, the same 100-turn workload fell from a median of 3.036 s per turn in early September to 414 ms; each turn sends one message, gets an immediate mock-model response and makes no tool calls, and timing excludes model generation. The framework hands users the loop that calls models, runs tools and delegates to subagents, with a resumable Workflow SDK built in.

Forecasting a virus's next move: antiGen shortlists mutations from the evolutionary tree, and Germinal Multi redesigns antibodies against them

Monoclonal antibodies were an important early tool against COVID-19, but they lost effect as SARS-CoV-2 mutated, and screening for antibodies that hold across variants struggles to keep up. antiGen learns from mutations already made along a virus's phylogenetic tree rather than from sequences alone, using roughly half a million parameters to forecast plausible next mutations; Germinal Multi tests one antibody design against several versions of a target at once on a computer. Combined, they redesigned BD55-1205, which neutralized XFG but not XFG-Next, into X51, which blocked infection by both in laboratory tests.

mastra.ai

Mastra exposes agents, workflows and tools through a built-in MCP server, letting coding agents call them directly

Mastra released a built-in MCP server: install @mastra/mcp and configure auth, and the whole Mastra server becomes available at /api/mcp to MCP clients. Tasks that previously had to be done by hand through the Studio UI can now be handled by configured coding agents or connected MCP clients. The server ships tools to call_agent, start_workflow and get_trace, and each call uses the same server routes as standard HTTP requests, passing the same RBAC and FGA checks.

Vercel News

Microsoft's MAI-Image-2.7 and Flash arrive on AI Gateway: one key, no Microsoft account, 2048×2048 output

Microsoft has placed MAI-Image-2.7 and MAI-Image-2.7-Flash on AI Gateway in public preview, and both raise the maximum output size to 2048 x 2048 pixels while adding transparent backgrounds for generated and edited images. A developer needs only one AI Gateway API key, with no Microsoft account, to generate images from text or edit existing images. Both models support photorealistic images, text rendering, and targeted edits to products, portraits, and layouts, and up to five reference images can guide edits or combine subjects.

OpenAI News

Astra for Law tops Harvey LAB-AA after update: 20.8% full credit on 120 legal assignments, 0.03 hallucinations each

OpenAI updated its legal model configuration Astra for Law and opened it for access requests to every ChatGPT Enterprise customer. The configuration builds on GPT-6.2 Astra and adds an expanded legal search index plus instructions for researching a question, weighing the authorities and writing the answer. On 120 Harvey LAB-AA assignments it earned full credit on 20.8%, above GPT-6.2 Astra at 15.0% and Muse Wonder Full at 13.1%, with 0.03 material hallucinations per assignment, tied for the lowest rate with GPT-6.

OpenAI News

OpenAI and Chile sign an MoU: no investment or purchasing commitment, with civil-servant AI training and startup support still to be scoped

OpenAI and the Government of Chile signed a Memorandum of Understanding launching OpenAI for Chile, led by Chile's Ministry of Science, Technology, Knowledge and Innovation, to support responsible AI adoption in public administration, build civil servants' skills and help local startups develop useful AI applications. The MoU was signed on the sidelines of Semafor's Silicon Valley & The World conference in California and is supported by the U.S. Department of Commerce's American SI Exports Program. A joint task force of ministry and OpenAI representatives is expected to agree priorities and an initial workplan, while specific activities, eligibility, resources and technology access will be agreed separately; the MoU sets no investment or purchasing commitment.

NVIDIA Blog

NVIDIA unveils Nemotron for Science at SC26: an agent at Argonne planned its own hypotheses and found two new binding sites

At SC26, NVIDIA announced Nemotron for Science, an open toolkit for science agents that combines the Nemotron 3.5 Super VL open model with a new open chemistry dataset and post-training recipes for life sciences and chemistry, runnable locally on a DGX Spark, a workstation or a researcher's own supercomputer. At Argonne National Laboratory, an agent built with the toolkit took a single prompt, planned and tested hypotheses, chose which simulations to run and identified two new binding sites. NVIDIA also announced the Rubin HPC GPU, an expanded Vera CPU platform and a quantum computing collaboration with IBM and Fermilab.

Microsoft bundles Copilot, agent governance and security models into one stack, with over 50 million agents already in Agent 365

At Ignite 2026 Microsoft combined Copilot, Microsoft IQ, Agent 365, the Perception security system and its own models into an enterprise "frontier ecosystem" stack, and declared Microsoft-Decision-1.1, the Code 2 family, Cyber 2 Flash and Perception generally available. Copilot is positioned as the entry point for work, folding Chat, Cowork, Code, Autopilot and Office together, with Work IQ underneath holding organizational context such as people, documents, email, meetings and projects. Agent 365 is a control plane that extends existing governance, identity, security and management systems to enterprise agents, and Microsoft says more than 50 million agents are already in it.

MongoDB | Blog

MongoDB buys Firebolt to put real-time analytics beside operational data, so agents query without pipelines

MongoDB has agreed to acquire analytics database company Firebolt and plans to integrate Firebolt's high-performance analytics engine into its Intelligent Data Platform, running alongside the operational database and kept in sync automatically so applications and AI agents can run complex queries on live operational data. Previously such analytics usually meant copying data into a separate system and maintaining pipelines and synchronization. MongoDB cites Similarweb running customer-facing analytics over more than a trillion rows at petabyte scale on Firebolt, with dynamic queries delivering sub-second performance.

Cradle

Cradle expands its antibody design platform to more argenx teams, where each experiment round sharpens the next design

Cradle announced that argenx, after successful results from initial work, is expanding use of Cradle's end-to-end AI-driven protein engineering platform to more scientists and antibody discovery programs. Previously argenx teams designed from their starting candidates; with the platform, scientists explored sequence space beyond those candidates across successive design, test, and learn rounds and ran designs independently, letting teams test more diverse ideas faster. The platform combines AI-driven engineering, active learning from experimental data, and multi-property optimization in one system, used to optimize existing leads and generate antibody candidates designed for properties including target affinity and selectivity.

Aignostics

HETA widens spatial H&E readouts to immune phenotyping and mitosis, ranking lymphocyte density in tumor core and margin against a pan-cancer TCGA reference

Aignostics updated HETA, its spatial H&E analysis platform, adding immune phenotyping readouts, tumor-infiltrating lymphocyte analysis, mitosis detection, and nerve and fat segmentation. Immune phenotyping measures lymphocyte density in the tumor core and at its margins separately, each ranked against a pan-cancer TCGA reference, as continuous values rather than one label per slide, so researchers can set their own inflamed, excluded and desert thresholds. The platform also outputs counts and densities of mature and immature tertiary lymphoid structures and extends tissue segmentation to 9 classes, as the basis for future perineural invasion readouts.

OpenAI News

OpenAI releases GPT-6.2 Astra and Sol: Astra tops its predecessor by 5.2 points on agentic professional work at under a third of the cost

OpenAI released GPT-6.2 Astra for the hardest projects and GPT-6.2 Sol for everyday work, both delivering better results on demanding tasks with fewer output tokens. On Agents' Last Exam, which tests multistep professional work, Astra scores 5.2 points above GPT-6 Astra at less than a third of the cost, and 5.5 points ahead of Claude Fable 5.2 at just 5% of the cost. Astra also scores 9 percentage points higher than GPT-6 Astra on Terminal-Bench-Science 0.1, which tests scientific work in a terminal, at less than one-third the cost.

Huawei Cloud open-sources CloudRobo-WM1.2, an embodied world model that outputs actions directly and tops two robot benchmarks

Huawei Cloud released and open-sourced CloudRobo-WM1.2 in Hangzhou, an Ascend-native embodied world model, aiming to move robots from recognizing static images to anticipating physical interactions and generating actions directly. Embodied intelligence has been held back by difficult multimodal physical understanding, tight coupling between world cognition and action generation, hard-to-scale heterogeneous data, weak cross-embodiment generalization and high real-time inference cost. The model uses a shared world-model backbone for prediction-enhanced world representation and a world-model-native DualAct architecture to output actions, ranking first on both RoboTwin 2.0 and LiberoPro 2.0.

bioRxiv

No structure needed first: Spec2Function reads metabolite function straight from tandem mass spectra, lifting explainable cirrhosis peaks past one quarter

Untargeted metabolomics has long required structural identification before pathway analysis, and incomplete spectral libraries leave many differential peaks outside downstream interpretation. Spec2Function aligns tandem mass spectra with multi-aspect functional text curated from HMDB in a shared space through dual-encoder contrastive learning, so a query spectrum needs no prior structural annotation at inference, and retrieval-augmented generation then produces literature-grounded functional summaries. On liver cirrhosis and a 30-organ macaque aging dataset it covered spectra beyond structural annotation and returned functional modules comparable across species.

langchain.com

LangSmith adds three access gates to production traces: engineers see run trees, not customer names or payment details

LangSmith has added three access controls to its Enterprise plans for production traces: admins can hide trace inputs and outputs, assign permissions by user group, and set trace-level policies from root run metadata. Previously teams either built custom observability setups with duplicated traces and multiple workspaces, which cost platform teams more to configure and maintain, or kept traces out of observability platforms, which left agent engineers without visibility. The controls layer together so engineers can inspect run trees, latency and cost while sensitive inputs and outputs stay with the people who need them.

bioRxiv

GatorAnchor scores each cell-cell link by cross-modal disagreement, topping mean clustering metrics on ten paired multi-omics datasets

In single-cell multi-omics integration, molecular layers often disagree about whether two cells are neighbors, yet most methods treat those relations as equally reliable. GatorAnchor folds each modality's uncertainty and the disagreement between modality means into a fused representation, then scores candidate cell relations from RNA similarity, second-modality similarity and shared edge-decoder support, weakening or removing unreliable edges. Across ten paired RNA+ADT and RNA+ATAC datasets it beat nine representative integration methods on mean AMI, NMI, ARI and ACC; keeping the 90% least-unreliable candidate edges raised same-type edge purity in all ten datasets, though graph structure changed differently by dataset.

bioRxiv

Loki2 reads molecular state from H&E at single-cell resolution, lifting cell-typing weighted F1 from 0.608 to 0.806

Routine H&E slides show cell morphology but not molecular state, and earlier pathology foundation models worked at multicellular resolution, leaving open whether a single cell's molecular program can be read from its appearance. Loki2 distills molecular supervision into morphology using UniSeg, a resource of about 15 million cells across 15 tissue types and 22 major lineages, aligning each nucleus's morphology embedding with a transcriptomic embedding. On cell typing it reaches weighted F1 0.806 and weighted accuracy 0.807, against weighted F1 between 0.574 and 0.608 for PLIP, GigaPath and UNI used as frozen feature extractors, and the genes it retrieves skew toward morphologically distinguishable cell identities and functions.

medRxiv

Letting the model learn its own target: LIFT-MASLD reaches a 0.836 predictive correlation in an external cohort

Prognostic models for metabolic dysfunction-associated steatotic liver disease (MASLD) usually fix an endpoint first and then improve the predictor. LIFT-MASLD reverses that order: SCORE learns a weighted phenotype from eight hepatic measures that is most predictable from baseline characteristics, with sparse regression selecting the baseline variables. The phenotype reached a predictive correlation of 0.860 in the HALO cohort, above 0.763 for predicting future liver stiffness from all baseline variables, and 0.836 in the NIDDK external cohort without refitting.

bioRxiv

FoldMapper breaks inverse folding's circular sequence search: shape-retrieved homologs lift best-of-100 recovery to 60.0% on TED

Protein inverse folding must design an amino-acid sequence that folds into a given backbone, yet using evolutionary information has required searching a multiple sequence alignment with the very sequence being designed, a circular dependency. FoldMapper instead retrieves structural homologs by shape with Foldseek and accumulates their amino-acid identities, weighted by bit score, into a per-residue structural position-specific frequency profile; a multiscale graph encoder then encodes that profile on a sparse local graph before a denser structure-conditioned message-passing network decodes the sequence. On TED structures it reaches 60.0% best-of-100 sequence recovery, above the 43.2% to 56.3% range of the other tabulated methods, with 20.1% sequence diversity.

medRxiv

Breast cancer five-year survival can be predicted without tissue imaging: HEAL-Surv lifts concordance from 0.658 to 0.704 by distilling a teacher

HEAL-Surv puts imaging mass cytometry (IMC, which measures protein markers per cell while keeping cell positions) and DNA methylation into a teacher used only during training, so a student predicts five-year net disease-specific survival in breast cancer from molecular and clinical inputs alone at inference. In five-fold cross-validation on 423 METABRIC patients, distillation raised the matched student's mean five-year Uno's concordance from 0.658 to 0.704, while the integrated and five-year Brier scores both rose.

bioRxiv

EnzRef reads reactants and products separately before touching the enzyme, lifting R² by 19.8% on unseen enzymes

Predicting enzyme turnover numbers (kcat) requires resolving activity differences between reactions of the same enzyme and between enzymes acting on the same reaction. EnzRef splits reaction chemistry into reactant and product conditions, interacts each separately with a frozen protein language model, expresses every conditioned enzyme readout as a deviation from that enzyme's own mean residue response, and recomposes the two responses. On the CatPred-DB source held-out set, mean single-model R² rises from 0.553 for MCKcat to 0.614; on the TurNuP-DB enzyme-unseen subset, R² rises from 0.259 to 0.310, a 19.8% relative gain without target-domain adaptation.

reflection.ai

Reflection opens its Trusted Access Program: autonomous pentest harness Laser goes first to vetted commercial partners

Reflection launched a Trusted Access Program that gives vetted organizations early access to selected AI capabilities, starting with Laser, an autonomous penetration-testing harness. Laser runs on Mirror, Reflection's terminal agent, and verifies findings with attack graphs and captured evidence rather than only generating plausible attacks. Laser is available as a research preview exclusively to commercial partners accepted into the program, for authorized security testing.

bioRxiv

Metabolites as nodes, reactions as hyperedges: HyperGenie predicts essential genes with 13% higher PR-AUC than objective-dependent methods

Predicting bacterial essential genes has relied on flux balance analysis, which assumes knockout strains still optimize for maximum growth, an assumption that does not always hold. HyperGenie represents the E. coli metabolic network as a directed hypergraph, with metabolites as nodes and reactions as hyperedges, and takes wild-type flux features directly as input, removing the post-knockout optimality assumption. On the iML1515 model under aerobic glucose, its PR-AUC is 13% above the objective-dependent FlowGAT and random forest, and it matches the flux-sampling FCL on comparable gene subsets.

bioRxiv

Even ideal artificial variants resist fine-tuning: LAMP shows sequence-to-function models still miss distal effects

LAMP injects artificial, strong and consistent transcription-factor-motif-perturbing SNVs into personal genome datasets to test whether sequence-to-function models can learn variant effects during fine-tuning. Even with the training set scaled to 12,111 genes and 12,011 such idealized causal SNVs, fine-tuned Enformer still trailed an elastic net baseline given only the SNV genotype on held-out genes and people. The analysis points to two obstacles: pre-trained priors that downweight distal regulatory elements make enhancer-to-gene links hard to strengthen, and sequence context differences between instances of the same motif obscure an otherwise easy-to-learn perturbation pattern.

bioRxiv

Moving CRISPRi knockdown responses to a new cell line takes a partial screen: at 2.5% screened, CASTER beats the transferred mean

Across CRISPRi Perturb-seq of essential genes in four human cell lines, the same knockdown's transcriptional response correlated at a median of 0.52 between lines after noise correction, so responses are only moderately shared. CASTER starts from a knockdown's mean response in the other lines and corrects it for the target line using part of that line's screened library; with 30% screened, a linear model and CASTER scored 0.84 and 0.80 on the calibrated scale versus 0.44 for the transferred mean. CASTER led only at 2.5% screened (0.62 versus 0.47), and with control cells alone and no screen it fell below the transferred mean.

bioRxiv

statLens splits statistical design among code, models and researchers, lifting small-model accuracy from 81.0% to 93.6%

Differential expression analysis requires settling the model family, primary comparison and sample dependence before fitting, decisions previously left to analysts or to code-writing autonomous agents. statLens assigns data profiling and rule checks to code, study-context interpretation to a language model, and final confirmation to the researcher; across 102 tasks, seven models held 92.5–95.0% accuracy, against 17.5–46.8% for LLM-only inference. Mean accuracy for four small models rose from 81.0% as autonomous agents to 93.6%.

bioRxiv

Protein-design pipelines learn to change course mid-run: IMPRESS reallocates sampling at runtime and solves five-island active sites a fixed cascade misses

Protein design usually chains diffusion backbone generation, inverse folding and structure prediction into a fixed pipeline whose sampling budget is set before the campaign starts, so filters can only discard candidates and cannot steer sampling toward regions the model stack can reach. IMPRESS makes each design pipeline a stateful object that halts, continues, resamples or branches at runtime from its own accumulated backbones, sequences and folds, with decisions evaluated asynchronously and without global synchronization. On the 41-motif AME enzyme-ligand benchmark, the adaptive workflow returns high-quality solutions for motifs with three or more discontinuous residue islands that the fixed cascade does not solve at all, including better than 40% of five-island motifs.

bioRxiv

SMSTPred types methyltransferase substrates from sequence alone: MCC 0.94 under homology-disjoint splits, ahead of sequence and structure homology search

SMSTPred is a multi-label classifier that reads the substrate type of SAM-dependent methyltransferases, DNA, RNA, protein or small-molecule metabolite, from protein language model embeddings without structural input. Under cross-validation with non-overlapping homology clusters it reaches a mean MCC of 0.94, above sequence-homology nearest neighbor MMseqs2 at 0.848, structure-homology Foldseek at 0.887 and a profile HMM at 0.519. Applied to 115,850 candidate methyltransferases from 1,249 reference proteomes, its predictions show a higher small-molecule share in plants and Actinomycetota, while the plant expansion mostly follows proteome size.

ICUBE Cognitive opens to conversational AI via a Tessera-powered MCP connector, letting compatible assistants query its tracked research

Molecular Connections has released a Tessera-powered MCP connector for its information-tracking platform ICUBE Cognitive, making the research, news and industry developments the platform gathers reachable through compatible AI applications. Previously that material was viewed inside the platform, with teams browsing and filtering it themselves; now organizations can explore relevant content through conversational AI. The company says this helps users find what they need faster and make better-informed decisions, but the announcement reports no usage, retrieval-speed or accuracy figures.

cuBLAS moves double-precision multiplication onto integer Tensor Cores: DGEMM up to 5.5x faster, CCSD 2x end to end

NVIDIA's cuBLAS in CUDA Toolkit 13.4.2 adds two algorithms, Ozaki-II and 2M, that split 64-bit floating-point matrix multiplication into integer pieces, multiply them on Tensor Cores and reassemble the exact result, speeding up DGEMM and ZGEMM with no application code changes. On VR NVL72 and GB200, Ozaki-II DGEMM runs 5.5x and 3.9x faster than native FP64, and a quantum-chemistry CCSD case is 2x and 1.4x faster end to end, with full-precision results indistinguishable from native FP64.

Gemini

Google releases Gemini 4 Hydrogen: a standalone coding model that also pairs with Carbon and Argon for agentic workflows

Google has released Gemini 4 Hydrogen for developers' everyday coding and agentic workflows. It runs standalone with four configurable thinking levels, or alongside frontier models such as Gemini 4 Carbon and Gemini 4 Argon, covering repository coding and debugging, interactive web development and multimodal workflows. Google says it follows instructions more tightly and makes fewer fabricated claims, and it is available through the Gemini API, Google AI Studio, Android Studio and Google Antigravity for paid subscribers.

MongoDB | Blog

MongoDB plugs a memory store into Vercel AI SDK: agents keep user context across sessions, with vector search in the same database

MongoDB released and open-sourced an integration that turns MongoDB Atlas into the persistent memory store for AI applications and agents built with the Vercel AI SDK. The SDK itself ships no database, so an application starts every conversation with no knowledge of the last one, and developers had to design schemas, write cleanup jobs and set up vector search themselves. The integration sorts memories into session, semantic, procedural, episodic and scratchpad tiers and stores them as Atlas documents, with the semantic, procedural and episodic tiers retrieved by meaning through MongoDB Vector Search rather than keywords, and indexes and embedding dimensions handled at startup.

AI

Google Search swaps in Gemini 4 Hydrogen: global users can now ask multi-part, typo-ridden questions with tool calls

Google has put Gemini 4 Hydrogen into Search for all users worldwide, calling it the biggest model upgrade to Search yet. Previously Search had limited grasp of multi-part questions, hard-to-read typos and highly specific requests; the new model deepens instruction following and tool calling and strengthens query fan-out, splitting a question into relevant subtopics searched at once to surface links users might not otherwise have found. Responses also lean more on images, structured content, comparison tables and custom tools.

Vercel News

Vercel ships Workflow SDK 5: run payloads are encrypted end to end by default, and a 1,020-step workflow's median finish drops from 21.0 to 3.0 minutes

Vercel released Workflow SDK 5, which encrypts durable workflow payloads end to end by default and keeps each run's state in the region where it started. Previously every run's event log, queue and streams lived only in iad1, so steps elsewhere made cross-region requests to record progress; v5 deploys Workflow infrastructure to every Vercel Function region and rewrites the engine to cut queue hops. On a 1,020-step benchmark, v5.2.0's median time to finish all steps fell from 21.0 minutes on v4.8.12 to 3.0 minutes.