Skip to main content
Search

University of California, Berkeley

31 related articles

Subscribe to updates
  1. arXiv

    Transluce automates the hunt for training data with surprising side effects, yielding 6,575 datasets that mostly transfer across models

    Researchers at Transluce and Berkeley turned datasets with surprising side effects from hand-built curiosities into something searchable at scale: given a target behavior, evolution, reinforcement learning and two subliminal-learning rewrites generate candidate data, and judges never used in the search filter it, producing 6,575 datasets across 8 behaviors. The data reads like ordinary post-training data yet makes a finetuned model show the target behavior; on 64 datasets, ML researchers predicted the side effect 63% of the time against 50% by chance.

  2. arXiv

    Route Anything extends routing to images, video, speech and robots: learned routers beat the best fixed candidate by 0.3 points across 16 benchmarks

    General-purpose AI systems are shifting from a single model to orchestrating many specialized models, tools and policies, but routing had mostly been studied on language models. Route Anything builds a unified trajectory matrix of 25 benchmarks, 452 candidate configurations and 374,401 scored runs, together with a routing theory for quality–cost trade-offs. On 19,814 test inputs the per-input oracle beats the best fixed candidate by 5.38 points; learned routers beat it by 0.3 points across 16 benchmarks, by 2.1 points on video generation, and by 1.1 points on robot manipulation with 24% less policy time.

  3. arXiv

    CyberArena puts AI agents in live attack-and-defense combat, and two frontier models pull ahead of the field

    CyberArena turns cybersecurity evaluation from static targets into multi-round combat: two agents in isolated sandboxes attack each other's services while patching their own, with objectives rotating every round. Across 40 challenges and 466 games, GPT-5.6 Sol and Claude Opus 5 led 12 frontier agents, and on 30 challenges all three compared models solved solo, round-three capture rates fell to 52%, 37% and 35%.

  4. arXiv

    Web agents stop deciding mid-page: Baya compiles the task first, cutting injection success to 0%

    Web agents usually follow ReAct, choosing each action from page observations that mix trusted and untrusted content, so injected text can rewrite their control flow. Baya compiles the user request into a Python program that calls a Trusted Website API before touching the live site; the planning model never sees live page content, and only the program reads it at runtime. On 581 WebArena Verified tasks Baya reaches 77.8% aggregate utility versus 55.1% for browser-only ReAct and 66.7% for API-enabled ReAct, and on WASP it records a 0% attack success rate against 16.2% and 25.5% for those baselines.

  5. arXiv

    Liver CT diagnosis switches to composing findings: diseases unseen in training are detected and localized

    Liver CT diagnosis models usually predict disease labels directly from images, so they only recognize diseases seen in training. DbC splits diagnosis into two steps: it learns a set of radiological findings as dense maps over the scan, then defines each disease by the findings its guidelines require and composes the diagnosis map with a fixed, parameter-free soft-minimum operator. On multi-phase liver CT, for six diseases never seen in training, DbC localized lesions significantly better than every comparator, including models that align images with disease names, and kept its advantage on rare diseases from another institution and on a public external cohort.

  6. arXiv

    Unit-test rewards teach models to edit more: PreciseCoder cuts average edits from 7–15 lines to 3 without losing pass rate

    Reinforcement learning from unit tests raises pass rates but also teaches code models to edit more and to invent bugs in programs that are already correct. PreciseCoder adds a signal for semantically verified edit precision and recall against a reference repair; on 4B, 9B and 27B models it cuts the average edit from 7–15 lines to 3 at a pass rate on par with unit-test-only training, and raises the no-edit rate on correct code to 47–52%.

  7. arXiv

    Attention now picks its own parameters per token: NESTOR nests expert routing inside column routing and lowers pretraining loss at equal compute

    Mixture-of-experts transformers apply conditional computation only to the MLP, so every token still passes through all attention parameters. NESTOR extends the same routing idea to attention: each token is sent to a few columns, each column has its own attention and experts, and the two levels of sparsity nest hierarchically. In language-model pretraining from 100M to 1B active parameters, the MoE transformer needs 1.13–1.29 times the compute and the dense transformer 3.1–4.1 times the compute to reach the same held-out loss, and this factor shows no downward trend as scale grows.

  8. Disordered proteins get all-atom ensembles in one diffusion pass, matching NMR and SAXS without reweighting

    IDPForge repurposes ESMFold's attention and structure modules as a diffusion generator that produces all-atom ensembles of intrinsically disordered proteins and regions directly from sequence, without per-sequence training, coarse-grained back-mapping or ensemble reweighting. On 32 test sequences with NMR and/or SAXS data it reaches a normalized total X-EISD score of 0.89, against 0.57 for IDPConformerGenerator, 0.58 for IDPFold and 0.58 for CALVADOS, and it can denoise only the disordered region around a folded template.

  9. arXiv

    Instruction-hierarchy training missed one conflict type: adding goal-goal pairs lifts Qwen3.5-9B conflict resistance by 34.3 points

    Existing instruction-hierarchy training sets cover constraint-constraint and constraint-goal conflicts but not two goals that conflict with each other. Decomposing a task into goals and constraints yields 9,544 goal-goal preference pairs built from two text-processing instructions that each process the other's message, and fine-tuning Qwen3.5-9B and Gemma4-12B-it on them with direct preference optimization raises mean conflict resistance by 34.3 and 19.4 points on four instruction-hierarchy benchmarks unseen in training, with small mean utility changes across eight benchmarks.

  10. arXiv

    Reasoning no longer blocks the action: masking current reasoning from the action expert keeps success and cuts latency up to 3.4x

    Reasoning vision-language-action models previously made every action wait for autoregressive decoding of the reasoning text, which accounted for 53% to 70% of per-call latency. TacitVLA masks the action expert's attention to the current reasoning tokens, keeping only the training effect of reasoning supervision on the backbone representation, and on memory tasks generates each ledger entry during action execution. Across RoboMME, RMBench, LIBERO and PhysicalAI-AV it matches its explicit-reasoning counterpart while cutting latency up to 3.4x, and matches no-reasoning latency while improving success by up to 17.1%.