AutoPass uses runtime and compiler evidence to guide LLM-generated optimization decisions, outperforming expert heuristics and classical autotuning methods. It achieves geometric-mean speedups of 1.043x on x86-64 and 1.117x on ARM64 systems without prior training or fine-tuning.
CRAX introduces a high-fidelity, accelerated safety benchmark for reinforcement learning using MuJoCo XLA. It achieves up to 100x speedups over CPU-based benchmarks via vectorization and hardware acceleration, featuring six environment suites and three agent-specific tasks across three difficulty levels. Evaluation of six safe RL methods shows no single approach dominates, highlighting trade-offs between performance and safety, with curriculum learning and safety transfer improving results.
VIMPO introduces a critic-free policy optimization method that derives a policy-implied value function from KL-regularized reinforcement learning. It enables verifiable reward incorporation without training a critic and outperforms GRPO on mathematical benchmarks, especially under noisy rewards.
A hierarchical system using a pretrained LLM to select RL skill policies outperforms flat RL in a 2v2 King of the Hill environment. It matches hand-crafted behavior tree performance in win rate and is perceived as more human-like by 60% of users, highlighting effective coordination and adaptability without manual rule design.
Pose6DAug enables robot data augmentation by swapping objects in successful episodes while preserving physically valid 6D pose trajectories. It operates in 3D using a mesh anchored by temporally coherent poses, ensuring multi-view consistency and physical plausibility. Fine-tuning a VLA policy on this augmented data improves novel object success rates by 16.5% over state-of-the-art baselines.
Benchmarks using fixed-shape checks miss real bugs in LLM-generated GPU kernels. A controlled corpus of 24 kernels, including 9 buggy variants with transcription errors, reveals that an op-schema-aware oracle detects all failures and passes all correct controls, with identical results across five GPU architectures.
CRAX introduces a high-fidelity, fast safety benchmark for reinforcement learning using MuJoCo XLA. It achieves up to 100x speedups over CPU-based benchmarks via vectorization and hardware acceleration, featuring six environment suites and three agent-specific tasks across three difficulty levels. Evaluation of six safe RL methods shows no single approach dominates, highlighting trade-offs between performance and safety, with curriculum learning and safety transfer improving results.
AgentFinVQA introduces a multi-agent pipeline for financial chart question answering that ensures auditability and on-premise deployability without significant accuracy loss. It outperforms baseline models by +7.68 pp using a proprietary backbone and +4.84 pp with open-weights Qwen3.6-27B-FP8, while providing a confidence signal via verifier output that improves human review routing.
Sevra, a serving-layer controller, selectively verifies answers to improve accuracy and reduce token usage. On \mathfive, it achieves 76.3% accuracy with 26.8% fewer post-generation tokens and halved harmful flips, while on \gsm it verifies only 3.0% of examples, boosting accuracy to 94.5% and cutting verification tokens by 91.2%. The study shows that initial solve length and explicit control needs determine optimal verification strategy.
A new framework defines when single-neuron interventions coherently control model behaviors without output collapse. The control window, based on alignment and norm ratios, predicts behavior triggers and collapse ceilings using forward pass data, with high accuracy on held-out neurons. On refusal, control is typed: coherent bypass occurs without actionable content, while genuine actionable reach appears only in specific cases and at later rollout stages.
AtomMem introduces a memory system that stores high-value atomic facts from long-form interactions. It uses hierarchical event structures and temporal profiles to capture coherent episodic contexts and track evolving user attributes, enabling stable and efficient memory evolution. Experiments on the LoCoMo benchmark show AtomMem achieves state-of-the-art performance in reasoning tasks.
GEMS enables training-free superposition of multiple semantic directions in LLMs by addressing distributional deviation and directional interference through geometric constraints. On GSM8K, it maintains 98% accuracy with three non-mathematical directions, while unconstrained addition drops to 4%; on Wikitext-2, it increases PPL by only 2.2%.
HydraHead introduces a head-level hybridization of Full and Linear Attention, leveraging interpretability to select retrieval-critical heads and fuse outputs via a scale-normalized module. Trained on 15B tokens, it achieves over 69% improvement over baseline at 512K context length, outperforming layer-wise hybrids and approaching Qwen3.5's performance on long-context tasks.
Fine-tuning language models on insecure code causes emergent misalignment. A shared activation direction across four model families achieves 99.6% separation of aligned and misaligned activations, and subtracting it reduces code spillover by 21-51 points. Cross-architecture transfer shows behavioral suppression but lacks specificity, with within-model directions being causally actionable and cross-model directions only causally real.
Discriminator-Guided RL (DRL) uses a pretrained representation space to train a discriminator that separates real data from model-generated samples. Its logit is used as a reward in KL-regularized RL, aligning model outputs with visual and semantic realism without human preferences. DRL improves FID and semantic FD across models like SiT and JiT, and enhances the Pareto frontier between preference and fidelity.
MAST, a mechanism-guided unlearning method, achieves targeted forgetting of RLVR-induced reasoning with minimal collateral damage. On Qwen2.5-Math-1.5B and Qwen3-1.7B-Base, it significantly reduces MATH performance (45/150 to 37/15-0) while preserving GSM8K accuracy by +0.8 points and maintaining MATH retention at -0.5 points. Results hold across different seeds, objectives, and models, showing superior stability over full-parameter unlearning.
STARE addresses policy entropy collapse in GRPO-based reinforcement learning by identifying entropy-critical token subsets via surprisal quantiles and reweighting their advantages. It maintains stable policy entropy across model scales and tasks, outperforming DAPO and other baselines by 4%-8% on AIME24 and AIME25, with consistent exploration-exploitation balance.
TxBench-PP is a verifiable benchmark for small-molecule preclinical pharmacology, testing AI agents' ability to derive accurate conclusions from real-world assay data. Across 16 model-harness configurations, no system reliably made correct preclinical pharmacology decisions, with the best performance at 59.3% (Claude Opus 4.8 / Pi) and 55.3% (GPT-5.5 / Pi) of endpoint attempts.
OneCanvas enables 3D scene understanding in Vision-Language Models by aggregating patch features onto a single panoramic canvas using 3D world coordinates. It achieves state-of-the-art performance on SQA3D and VSI-Bench, and generalizes to out-of-distribution data on SPBench, using significantly less training compute than existing methods.
Diffusion-Proof is the first framework to train and apply diffusion language models for formal theorem proving. It introduces dLLM-Prover-7B for whole-proof writing with long-range coherence and dLLM-Corrector-7- for local proof correction using bidirectional information. The framework outperforms auto-regressive LLM baselines by 1.61% on ProofNet-Test and 6.14% on MiniF2F-Test, and solves an IMO problem beyond the capability of DeepSeek-Prover-V2-7B.