Skip to main content

AI Core

  1. 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.

  2. 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.

  3. TokenEconomizer lets coding agents learn from past sessions: 18.4%–30.9% lower cost across 277 real bug-fixing tasks

    Developers use AI coding agents in the same codebases day after day, yet agents often start each session from scratch, rediscovering familiar facts, loading unused tools and repeating unsuccessful approaches. TokenEconomizer, an open-source plugin from UC Berkeley, Microsoft Research and Microsoft Security, turns completed session logs into guidance for later sessions without modifying the agent or its underlying model. Across 277 real bug-fixing tasks in six repositories, it cut average cost by 18.4%–30.9% for five agent and model combinations, saving in all 30 repository-combination pairs while task success rates stayed broadly stable.

  4. 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.

  5. 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.

  6. 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.

  7. 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.

  8. 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%.

  9. 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.

  10. arXiv

    Latent feedback makes each FLOP go further: up to 48% less compute, or 47% less unique data, than repeating epochs

    A latent-feedback Transformer feeds each token's top-layer state into the next token's input; earlier comparisons held training tokens fixed and never repeated data. Across 60M to 778M parameters, 12.5 to 400 unique tokens per non-embedding parameter and one to four epochs, with both models tuned at every size and data amount, latent feedback reaches a Transformer's loss with less compute at fixed unique data and with less unique data at fixed epochs, and both savings grow with unique data per parameter and with repetition. A fitted scaling law predicts held-out 778M losses to within 0.3% and the gap between the two models to within 0.05%.