30
30
30
30
30
Coupled Calibration and Learning: Mitigating Teacher Bias in LLM Distillation without Target-Domain Reward Feedback
CCL couples teacher calibration with student updates via token-level branching, provably removing teacher bias in LLM distillation.
The paper proposes Coupled Calibration and Learning (CCL), an LLM distillation algorithm that alternates teacher calibration using source-question reward feedback with student training on target questions under covariate shift. Each iteration calibrates the teacher on source feedback, trains the student on target questions, and lets the updated student inform subsequent calibration. The authors prove the student's expected KL divergence to the oracle student converges to zero at a polynomial rate, and show regularized direct matching error can remain bounded away from zero.
20
Mind2Dialogue: Training Human-Aware Language Models by Simulating User Mental States
Mind2Dialogue simulates users' mental states to generate privileged supervision, boosting personalization and preference-following in Qwen, Llama, and OLMo assistants.
The Mind2Dialogue framework uses a psychology-guided simulator that preserves personal characteristics while updating user mental states through interaction, driving coherent conversations and an Oracle assistant's responses. Privileged distillation trains models on the Oracle's well-informed responses so they can assist users without direct access to mental states at deployment. Training on the full corpus improves every reported personalization metric over Qwen, Llama, and OLMo instruction-tuned baselines, including 26.6 to 40.9 percentage point gains in preference-following generation.
45
Beyond Solver Verdicts: Generative Reward Models for Autoformalization
Researchers introduce Generative Verification (GenV), a generative reward model achieving 0.961 AUROC in detecting unfaithful autoformalization that preserves solver verdicts.
The paper formalizes Verdict-Preserving-Unfaithfulness (VPU), a failure mode in neurosymbolic autoformalization where an incorrect encoding executes successfully and matches the expected solver verdict, and proves verdict-only verification is bounded to chance-level detection. The proposed Generative Verification (GenV) distills an offline Z3-equivalence oracle into a reference-free, continuous reference-equivalence score within the language model's vocabulary space. The oracle-mined verifier (GenV+HN) achieves 0.961 AUROC, generalizes zero-shot across unseen translators and formal styles, and yields an 11.3-point downstream accuracy gain in agentic test-time compute allocation. Mechanistic analysis with decision-projected logit lenses and sparse autoencoders shows the generative readout extracts precise spatial error coordinates without explicit localization training.
35
Meta FAIR Introduces AI Research Preference Models (RPMs): Ranking ML Experiments Before Spending GPU Hours
Meta FAIR, Oxford and UCL introduce Research Preference Models that rank unexecuted ML experiments, lifting AIRS-Bench scores from 0.684 to 0.729 and cutting compute ~1.6×.
Researchers from Meta FAIR, Oxford, and UCL introduce Research Preference Models (RPMs), which use frozen pretrained LLMs (Qwen3.6-27B backbone, no fine-tuning) to rank unexecuted experiment candidates and execute only the winner of a pairwise knockout tournament. Two variants shipped: an inference-only LLM-as-a-judge and an agentic variant that runs small pilot experiments in an H200 sandbox. On AIRS-Bench (20 tasks, 24 hours on one H200, 10 seeds), scores rise from 0.684 (random) to 0.711 and 0.729 versus a 0.748 validation oracle, and both variants reach the baseline's 24-hour score in roughly 15 hours. The team reports new SOTA on WinoGrande (94.1% with Agentic RPM) and SVAMP (95.7% with inference-only).
45
30
30
ImpossibleRubrics: Stress-Testing Generated Rubrics as Reward Signals
ImpossibleRubrics benchmark shows LLM-generated rubric reward signals are exploited 8-26% of the time by adversarial answers on impossible tasks.
ImpossibleRubrics is a benchmark of 169 impossible tasks across six impossibility categories, each paired with a verifiable oracle certificate, plus 48 answerable controls, for stress-testing LLM-generated rubrics used as reward signals. Eleven rubric generators were exploited 8-26% of the time on an unbiased 150-task cut and up to 36% on a stress cut, while a certificate-faithful rubric scored 0%. A single generic 'be decisive, penalize hedging' rubric was exploited 64% of the time, suggesting tailored criteria can reveal which claims attackers should fabricate.
45
E2A-Bench: Benchmarking Evidence-to-Action Reliability in Financial Chart Reasoning
E2A-Bench, a 969-query financial chart reasoning benchmark, finds VLMs fail evidence-to-action consistency, with fine-tuning amplifying BUY:SELL bias 4-6x.
E2A-Bench is a 969-query benchmark built from 323 HS300 constituents across three input modalities with deterministic OHLCV-derived evidence anchors, evaluating grounding, reasoning-action consistency, evidence-confidence calibration, and directional coverage via UCR, RCI, ECI, and NDR metrics. Testing 20 VLMs showed the lowest-hallucination model ranked near the bottom on coverage with only 6.4% directional coverage, and oracle-aided verification reduced unsupported claims but could collapse coverage. Financial fine-tuning amplified the BUY:SELL ratio by factors of 4.21 to 4.68 across base-fine-tuned pairs.
34
IdeaAMBIG: Benchmarking Implementation-Critical Gaps in Research-Idea Specifications
IdeaAMBIG benchmark with 660 instances measures whether LLMs can spot and fix underspecified research-method details for faithful implementation.
Researchers introduce IdeaAMBIG, a benchmark of 660 evidence-grounded instances (163 real-world gaps from reproducibility reports and GitHub issues, 497 controlled synthetic gaps) built from papers, codebases, and reproduction artifacts. It evaluates codification-readiness assessment, defect localization, and clarification action generation. Across 13 LLMs, the best model achieved only 9.6% Macro Defect Recovery Rate on real-world instances but 80.6% clarification success when given the annotated defect. An oracle study showed gold resolutions raise the codification-ready rate from 14% to 98%, identifying defect localization as the main bottleneck.
28
IdeaAMBIG: Benchmarking Implementation-Critical Gaps in Research-Idea Specifications
IdeaAMBIG benchmark of 660 specification-gap instances shows LLMs localize implementation-critical research gaps poorly, with best model at 9.6% defect recovery.
IdeaAMBIG is a benchmark of 660 evidence-grounded instances evaluating whether research-method specifications provide enough information for faithful implementation: 163 real-world gaps from reproducibility reports and GitHub issues plus 497 controlled synthetic gaps. It tests codification-readiness assessment, defect localization, and clarification action generation across 13 LLMs. The best model achieves only a 9.6% Macro Defect Recovery Rate on real-world instances, though 80.6% clarification success when given the annotated defect, and an oracle study shows gold resolutions raise codification-ready rates from 14% to 98%. Defect localization emerges as the main bottleneck across all evaluated models.
30
ExecCritic: Learn to Test, Test to Improve for Coding Agents
ExecCritic separates test generation from patching for coding agents, lifting SWE-bench Verified resolution to 72.6%.
ExecCritic pairs a test-verify-revise scaffold with role-specific reinforcement learning: a Test agent writes repository-native tests and a Repair agent fixes code from execution feedback, both using Qwen-3.5-35B-A3B backbones. Post-trained Qwen agents compose to 72.6% on SWE-bench Verified, an 11.4-point gain over the 61.2% no-test baseline, without stronger-model or oracle feedback at evaluation time. The work shows test quality is the key variable: base-agent tests lowered resolution to 57.3% while GPT-5.6-sol tests raised it to 65.3%.
36
The Surprising Effectiveness of Approximate Value Iteration in Self-Play
Minimal approximate value iteration self-play learns more accurate value functions than AlphaZero in Connect Four and Hex while cutting training and inference costs.
The paper trains a minimal self-play implementation of Approximate Value Iteration (AVI) without MCTS and uses ground-truth oracles for exact evaluation in Connect Four, 7x7 Hex, and synthetic games. AVI learns more accurate value functions than AlphaZero, and its one-step-lookahead greedy policies remain competitive with MCTS-based policies at substantially lower training and inference cost. Preliminary experiments on Othello and 9x9 Go show AVI trains stably on larger games, suggesting simpler approaches have become increasingly practical with modern deep-learning tools.
35
30
30
30
30
30