{"total":128,"items":[{"citing_arxiv_id":"2606.32029","ref_index":4,"ref_count":1,"confidence":0.98,"is_internal_anchor":true,"paper_title":"When LLMs Read Tables Carelessly: Measuring and Reducing Data Referencing Errors","primary_cat":"cs.CL","submitted_at":"2026-06-30T17:54:50+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":6.0,"formal_verification":"none","one_line_summary":"LLMs exhibit data referencing errors across model sizes; a critic model detects them at 78.2% F1 and boosts accuracy up to 12% via filtering and rejection sampling.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2606.31779","ref_index":58,"ref_count":1,"confidence":0.98,"is_internal_anchor":true,"paper_title":"Bridging the Gap Between Latent and Explicit Reasoning with Looped Transformers","primary_cat":"cs.LG","submitted_at":"2026-06-30T14:58:53+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":6.0,"formal_verification":"none","one_line_summary":"LOTUS uses a looped padded Transformer with parallel cross-entropy supervision on gold CoT tokens to match explicit CoT performance at 3B parameters while reducing thought-phase latency 2.5x-6.9x.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2606.31748","ref_index":12,"ref_count":1,"confidence":0.98,"is_internal_anchor":true,"paper_title":"Addressing Over-Refusal in LLMs with Competing Rewards","primary_cat":"cs.LG","submitted_at":"2026-06-30T14:38:49+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":6.0,"formal_verification":"none","one_line_summary":"SEAR trains one LLM via adversarial process rewards to explore harmful reasoning paths but flip to safe outputs, reducing over-refusal while preserving safety.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2606.30345","ref_index":11,"ref_count":1,"confidence":0.98,"is_internal_anchor":true,"paper_title":"DRIFT: Difficulty Routing Self-DIstillation with Rhythm-Gated Exploration and Success BuFfer Training","primary_cat":"cs.LG","submitted_at":"2026-06-29T14:20:47+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":5.0,"formal_verification":"none","one_line_summary":"DRIFT is an online self-evolution policy optimization framework using Difficulty Routing, Rhythm Gating, success buffers, and two-stage curriculum learning that reports new SOTA results on five reasoning benchmarks.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2606.30128","ref_index":9,"ref_count":1,"confidence":0.98,"is_internal_anchor":true,"paper_title":"Does Verbose Chain-of-Thought Really Help? In-Distribution Evidence that Content, Not Length, Matters","primary_cat":"cs.AI","submitted_at":"2026-06-29T11:06:46+00:00","verdict":"ACCEPT","verdict_confidence":"MODERATE","novelty_score":7.0,"formal_verification":"none","one_line_summary":"In-distribution sampling across 25 models and controlled interventions with DAG-verified content show that semantic reasoning and validation content, not token count, drive CoT gains.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2606.23364","ref_index":34,"ref_count":1,"confidence":0.98,"is_internal_anchor":true,"paper_title":"Convergence of Gradient Descent for General Neural Network Architectures Beyond the NTK Regime","primary_cat":"cs.LG","submitted_at":"2026-06-22T14:00:26+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":7.0,"formal_verification":"none","one_line_summary":"Proves GD convergence to stationary point neighborhoods for general NN architectures beyond NTK via block-level analysis, analyticity, and local smoothness conditions.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2606.22317","ref_index":22,"ref_count":1,"confidence":0.98,"is_internal_anchor":true,"paper_title":"Curriculum Reinforcement Learning Can Incentivize Reasoning Capacity in LLMs Beyond the Base Model","primary_cat":"cs.LG","submitted_at":"2026-06-21T03:15:25+00:00","verdict":"UNVERDICTED","verdict_confidence":"MODERATE","novelty_score":6.0,"formal_verification":"none","one_line_summary":"Boundary-aware Curriculum RL raises average pass@256 by 9.8 points over base models and 10.3 points over vanilla RLVR on Qwen, Llama, and DeepSeek families.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2606.21884","ref_index":23,"ref_count":1,"confidence":0.98,"is_internal_anchor":true,"paper_title":"A Verifiable Search Is Not a Learnable Chain-of-Thought","primary_cat":"cs.LG","submitted_at":"2026-06-20T04:52:14+00:00","verdict":"UNVERDICTED","verdict_confidence":"MODERATE","novelty_score":7.0,"formal_verification":"none","one_line_summary":"Verifiable search procedures cannot be learned as forward chain-of-thought by language models; they instead learn memorization, verification, or require precomputed catalogs.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2606.21228","ref_index":168,"ref_count":1,"confidence":0.98,"is_internal_anchor":true,"paper_title":"Sakana Fugu Technical Report","primary_cat":"cs.LG","submitted_at":"2026-06-19T08:47:40+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":5.0,"formal_verification":"none","one_line_summary":"Sakana Fugu trains LLM orchestrators using fine-tuning, evolutionary algorithms, and RL to build query-adaptive multi-agent scaffolds, claiming SOTA results on benchmarks including SWE-Bench Pro and GPQA-Diamond.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2606.17890","ref_index":21,"ref_count":1,"confidence":0.98,"is_internal_anchor":true,"paper_title":"Dynamic Rollout Editing for Reducing Overthinking in RL-Trained Reasoning Models","primary_cat":"cs.CL","submitted_at":"2026-06-16T13:10:30+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":6.0,"formal_verification":"none","one_line_summary":"Dynamic Rollout Editing reduces overthinking in RL-trained LLMs by editing post-answer continuations in successful rollouts and preferring the edited versions within GRPO groups.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2606.17687","ref_index":78,"ref_count":1,"confidence":0.98,"is_internal_anchor":true,"paper_title":"SuCo: Sufficiency-guided Continuous Adaptive Reasoning","primary_cat":"cs.CL","submitted_at":"2026-06-16T08:52:15+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":4.0,"formal_verification":"none","one_line_summary":"SuCo defines minimal sufficient CoT and applies a two-stage fine-tuning plus RL framework to enable continuous adaptive reasoning control, claiming gains in both accuracy and token efficiency on math, code, and science tasks.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2606.17539","ref_index":89,"ref_count":1,"confidence":0.98,"is_internal_anchor":true,"paper_title":"Reinforcing Dual-Path Reasoning in Spatial Vision Language Models","primary_cat":"cs.CV","submitted_at":"2026-06-16T05:32:39+00:00","verdict":"UNVERDICTED","verdict_confidence":"UNKNOWN","novelty_score":6.0,"formal_verification":"none","one_line_summary":"SR-REAL equips spatial VLMs with dual LOR and DTR reasoning paths trained via RL, achieving better benchmark performance through mutual reinforcement and generalization without per-task tuning.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2606.12578","ref_index":47,"ref_count":1,"confidence":0.98,"is_internal_anchor":true,"paper_title":"MARD: Mirror-Augmented Reasoning Distillation for Mechanism-Level Drug-Drug Interaction Prediction","primary_cat":"cs.CL","submitted_at":"2026-06-10T18:26:11+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":6.0,"formal_verification":"none","one_line_summary":"MARD-7B outperforms baselines and GPT-4o on novel drug pairs for mechanism-level DDI prediction via a new distillation pipeline with verifiable process rewards and releases all resources.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2606.10403","ref_index":11,"ref_count":1,"confidence":0.98,"is_internal_anchor":true,"paper_title":"KCSAT-ML: Probing Reasoning Models with Nationwide-Cohort Human Difficulty","primary_cat":"cs.CL","submitted_at":"2026-06-09T04:25:44+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":7.0,"formal_verification":"none","one_line_summary":"KCSAT-ML benchmark supplies human error rates for math problems and DRG metric exposes that model accuracy collapses on high-human-error items while test-time scaling shows non-monotonic gains and alignment failures.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2606.08728","ref_index":61,"ref_count":1,"confidence":0.98,"is_internal_anchor":true,"paper_title":"Artificial Intelligence for Mathematical Reasoning: An Integrated Survey of Language Models, Neuro-symbolic Systems, and Verified Discovery","primary_cat":"cs.AI","submitted_at":"2026-06-07T16:50:07+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":3.0,"formal_verification":"none","one_line_summary":"An integrated survey organizing AI mathematical reasoning into informal, formal, discovery, and technique axes while cataloging benchmarks and assessing failure modes.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2606.08501","ref_index":52,"ref_count":1,"confidence":0.98,"is_internal_anchor":true,"paper_title":"Back on Track: Aligning Rewards and States for Reasoning in Diffusion Large Language Models","primary_cat":"cs.CL","submitted_at":"2026-06-07T07:59:55+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":6.0,"formal_verification":"none","one_line_summary":"PAPO improves reasoning performance in diffusion LLMs by converting sparse terminal rewards into dense step-wise credit and replaying real high-uncertainty trajectories, reporting gains up to 42.2% on Countdown.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2606.08098","ref_index":57,"ref_count":1,"confidence":0.98,"is_internal_anchor":true,"paper_title":"When Does Delegation Beat Majority? A Delegation-Based Aggregator for Multi-Sample LLM Inference","primary_cat":"cs.AI","submitted_at":"2026-06-06T10:57:45+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":5.0,"formal_verification":"none","one_line_summary":"PPV delegation using letter entropy and per-question embedding cosine beats majority voting by 1.5 pp overall on MMLU-Pro in an unsupervised setting.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2606.06745","ref_index":28,"ref_count":1,"confidence":0.98,"is_internal_anchor":true,"paper_title":"When to Think Deeply: Inhibitory Deliberation for LLM Reasoning","primary_cat":"cs.CL","submitted_at":"2026-06-04T21:57:27+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":7.0,"formal_verification":"none","one_line_summary":"IDPR is a response-conditioned inhibitory deliberation method that trains a controller on fast-slow outcome pairs to decide when to override LLM fast answers, improving accuracy from 47.90% to 48.92% with slow reasoning invoked on only 8.20% of a 5,000-example math test set.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2606.05145","ref_index":21,"ref_count":1,"confidence":0.98,"is_internal_anchor":true,"paper_title":"Failed Reasoning Traces Tell You What Is Fixable (But Not by Reading Them)","primary_cat":"cs.LG","submitted_at":"2026-06-03T17:50:26+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":7.0,"formal_verification":"none","one_line_summary":"Three problem-level trajectory features derived from the distributional signature of failed LLM rollouts enable failure clustering at 84.3% accuracy and a training-free routing rule that improves rescue by 12.2% on hard cases.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2606.02248","ref_index":9,"ref_count":1,"confidence":0.98,"is_internal_anchor":true,"paper_title":"Geometric Latent Reasoning Induces Shorter Generations in LLMs","primary_cat":"cs.CL","submitted_at":"2026-06-01T13:40:55+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":6.0,"formal_verification":"none","one_line_summary":"GLR formulates latent reasoning as geometric path approximation in pretrained embedding space and reports shorter LLM generations on math tasks without an explicit length penalty.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2605.28774","ref_index":3,"ref_count":1,"confidence":0.98,"is_internal_anchor":true,"paper_title":"Agent Explorative Policy Optimization for Multimodal Agentic Reasoning","primary_cat":"cs.CL","submitted_at":"2026-05-27T17:36:39+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":6.0,"formal_verification":"none","one_line_summary":"AXPO addresses the Thinking-Acting Gap in agentic RL training by targeted resampling of tool calls in all-wrong subgroups, delivering +1.8pp gains over GRPO on nine multimodal benchmarks with an 8B model beating a 32B baseline on Pass@4.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2605.28070","ref_index":34,"ref_count":1,"confidence":0.98,"is_internal_anchor":true,"paper_title":"Bridging the Detection-to-Abstention Gap in Reasoning Models under Insufficient Information","primary_cat":"cs.AI","submitted_at":"2026-05-27T07:28:25+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":5.0,"formal_verification":"none","one_line_summary":"JTS trains reasoning models via supervised warm-up and missing-premise RL to make an explicit answerability commitment that triggers early termination on unanswerable inputs, raising Abstention@Detection near saturation.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2605.25183","ref_index":43,"ref_count":1,"confidence":0.98,"is_internal_anchor":true,"paper_title":"Knowledge Graph-Driven Expert-Level Reasoning for Neuroscience","primary_cat":"cs.CL","submitted_at":"2026-05-24T17:23:41+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":4.0,"formal_verification":"none","one_line_summary":"A textbook-derived neuroscience knowledge graph supplies synthetic multi-hop QA supervision and RL rewards to fine-tune a small LM claimed to exceed larger general models on expert reasoning.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2605.24785","ref_index":8,"ref_count":1,"confidence":0.98,"is_internal_anchor":true,"paper_title":"PANDO: Efficient Multimodal AI Agents via Online Skill Distillation","primary_cat":"cs.AI","submitted_at":"2026-05-24T00:07:25+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":7.0,"formal_verification":"none","one_line_summary":"PANDO introduces an online skill-distillation method with a structured library, reflection, demotion, routing, compression, and cache-aware prompting that reaches 58.3% success on 910 VisualWebArena tasks using 58-61% fewer tokens than prior methods.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2605.23074","ref_index":42,"ref_count":1,"confidence":0.98,"is_internal_anchor":true,"paper_title":"PathCal: State-Aware Reflection-Marker Calibration for Efficient Reasoning","primary_cat":"cs.AI","submitted_at":"2026-05-21T22:13:20+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":6.0,"formal_verification":"none","one_line_summary":"PathCal calibrates reasoning paths by type-aware soft rebalancing of reflection-marker logits at uncertain states, yielding better efficiency-performance trade-offs on six benchmarks.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2605.22064","ref_index":57,"ref_count":1,"confidence":0.98,"is_internal_anchor":true,"paper_title":"Hy-MT2: A Family of Fast, Efficient and Powerful Multilingual Translation Models in the Wild","primary_cat":"cs.CL","submitted_at":"2026-05-21T07:00:06+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":3.0,"formal_verification":"none","one_line_summary":"Hy-MT2 presents three new multilingual translation models that claim to outperform listed open-source and commercial systems on diverse tasks while enabling low-storage on-device use.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2605.20863","ref_index":19,"ref_count":1,"confidence":0.98,"is_internal_anchor":true,"paper_title":"PlexRL: Cluster-Level Orchestration of Serviceized LLM Execution for RLVR","primary_cat":"cs.DC","submitted_at":"2026-05-20T07:55:06+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":5.0,"formal_verification":"none","one_line_summary":"PlexRL multiplexes unified LLM services across RLVR jobs at the cluster level to exploit anti-correlated idle times and reduce GPU-hour costs by up to 37.58% with minimal per-job overhead.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2605.19358","ref_index":13,"ref_count":1,"confidence":0.98,"is_internal_anchor":true,"paper_title":"Taming the Thinker: Conditional Entropy Shaping for Adaptive LLM Reasoning","primary_cat":"cs.CL","submitted_at":"2026-05-19T04:41:51+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":6.0,"formal_verification":"none","one_line_summary":"CES applies conditional bidirectional entropy control on top of DAPO to improve accuracy and shorten responses on mathematical benchmarks for 7B and 1.5B LLMs.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2605.19285","ref_index":24,"ref_count":1,"confidence":0.98,"is_internal_anchor":true,"paper_title":"Are Rationales Necessary and Sufficient? Tuning LLMs for Explainable Misinformation Detection","primary_cat":"cs.CL","submitted_at":"2026-05-19T03:03:32+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":6.0,"formal_verification":"none","one_line_summary":"LONSREX introduces a metric-based pipeline to identify necessary and sufficient rationales when creating training data for fine-tuning LLMs on explainable misinformation detection, addressing limitations of naive label-based filtering.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2605.17497","ref_index":46,"ref_count":1,"confidence":0.98,"is_internal_anchor":true,"paper_title":"Self-Supervised On-Policy Distillation for Reasoning Language Models","primary_cat":"cs.LG","submitted_at":"2026-05-17T15:14:24+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":6.0,"formal_verification":"none","one_line_summary":"SSOPD converts intra-group correct-wrong contrast into process supervision by distilling a teacher distribution from the shortest correct completion into prefixes of the longest wrong completion, improving GRPO on AIME and HMMT benchmarks.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2605.17187","ref_index":184,"ref_count":1,"confidence":0.98,"is_internal_anchor":true,"paper_title":"PluRule: A Benchmark for Moderating Pluralistic Communities on Social Media","primary_cat":"cs.CL","submitted_at":"2026-05-16T22:52:11+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":7.0,"formal_verification":"none","one_line_summary":"PluRule is a new multimodal multilingual benchmark showing that state-of-the-art vision-language models perform only marginally better than a trivial baseline at detecting specific rule violations in pluralistic online communities.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2605.16826","ref_index":30,"ref_count":1,"confidence":0.98,"is_internal_anchor":true,"paper_title":"Decoupling KL and Trajectories: A Unified Perspective for SFT, DAgger, Offline RL, and OPD in LLM Distillation","primary_cat":"cs.LG","submitted_at":"2026-05-16T06:05:27+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":7.0,"formal_verification":"none","one_line_summary":"Decoupling prefix source from token-level KL direction in autoregressive sequence KL yields four objectives unifying SFT, DAgger, offline RL and OPD, with KL mixing and entropy-gated curriculum improving math reasoning accuracy and shortening responses.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2605.14358","ref_index":27,"ref_count":1,"confidence":0.98,"is_internal_anchor":true,"paper_title":"Uncovering the Representation Geometry of Minimal Cores in Overcomplete Reasoning Traces","primary_cat":"cs.AI","submitted_at":"2026-05-14T04:35:45+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":7.0,"formal_verification":"none","one_line_summary":"Language models produce overcomplete reasoning traces where on average 46% of steps can be removed while preserving the answer in 86% of cases, with necessity concentrated in the top three steps.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2605.14186","ref_index":32,"ref_count":1,"confidence":0.98,"is_internal_anchor":true,"paper_title":"LLMs Know When They Know, but Do Not Act on It: A Metacognitive Harness for Test-time Scaling","primary_cat":"cs.LG","submitted_at":"2026-05-13T23:09:25+00:00","verdict":"CONDITIONAL","verdict_confidence":"LOW","novelty_score":6.0,"formal_verification":"none","one_line_summary":"A metacognitive harness uses LLMs' pre- and post-solution self-monitoring signals to control test-time reasoning, raising pooled accuracy from 48.3% to 56.9% on text, code, and multimodal benchmarks.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2605.12813","ref_index":115,"ref_count":1,"confidence":0.98,"is_internal_anchor":true,"paper_title":"REALISTA: Realistic Latent Adversarial Attacks that Elicit LLM Hallucinations","primary_cat":"cs.CL","submitted_at":"2026-05-12T23:13:50+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":6.0,"formal_verification":"none","one_line_summary":"REALISTA generates semantically coherent adversarial prompts via latent-space optimization over input-dependent editing directions, achieving stronger hallucination elicitation than prior realistic attacks on open-source and reasoning LLMs.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2605.12652","ref_index":33,"ref_count":1,"confidence":0.98,"is_internal_anchor":true,"paper_title":"Multi-Rollout On-Policy Distillation via Peer Successes and Failures","primary_cat":"cs.LG","submitted_at":"2026-05-12T18:57:44+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":6.0,"formal_verification":"none","one_line_summary":"MOPD improves on-policy distillation by using peer successes and failures from multiple rollouts to construct more informative teacher signals, yielding consistent gains over baselines on reasoning benchmarks.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2605.11538","ref_index":6,"ref_count":1,"confidence":0.98,"is_internal_anchor":true,"paper_title":"Taming Extreme Tokens: Covariance-Aware GRPO with Gaussian-Kernel Advantage Reweighting","primary_cat":"cs.CL","submitted_at":"2026-05-12T05:05:47+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":6.0,"formal_verification":"none","one_line_summary":"Covariance-weighted GRPO with Gaussian-kernel reweighting tames extreme tokens to stabilize training and boost reasoning performance over standard GRPO.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2605.10930","ref_index":50,"ref_count":2,"confidence":0.98,"is_internal_anchor":true,"paper_title":"Evaluating the False Trust Engendered by LLM Explanations","primary_cat":"cs.HC","submitted_at":"2026-05-11T17:58:12+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":5.0,"formal_verification":"none","one_line_summary":"LLM reasoning traces and post-hoc explanations increase false trust in incorrect predictions, whereas contrastive dual explanations enhance users' ability to distinguish correct from incorrect AI outputs.","context_count":1,"top_context_role":"background","top_context_polarity":"background","context_text":"LLM generated explanations excel at this, as they are optimized to produce responses that are helpful, warm, and satisfying to users. Recent work has shown that such optimization can increase sycophancy and reduce the accuracy of the model i.e. the tendency to produce responses that agree with or please the user rather than responses that are correct [50, 51]. Models generating post-hoc explanations may be exhibiting similar tendencies. Dual side explanations help users to calibrate their trust:Motivated by the observation that one-sided explanations consistently increase false trust, we tested a dual explanation paradigm that presents both the merits and demerits of the AI's answer. The intuition is analogous to a peer"},{"citing_arxiv_id":"2605.08975","ref_index":36,"ref_count":1,"confidence":0.98,"is_internal_anchor":true,"paper_title":"Latency Analysis and Optimization of Alpamayo 1 via Efficient Trajectory Generation","primary_cat":"cs.AI","submitted_at":"2026-05-09T14:34:00+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":5.0,"formal_verification":"none","one_line_summary":"Redesigning Alpamayo 1 to single-reasoning and optimizing diffusion action generation cuts inference latency by 69.23% while preserving trajectory diversity and prediction quality.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2605.08441","ref_index":21,"ref_count":1,"confidence":0.98,"is_internal_anchor":true,"paper_title":"DUET: Optimize Token-Budget Allocation for Reinforcement Learning with Verifiable Rewards","primary_cat":"cs.LG","submitted_at":"2026-05-08T20:03:19+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":7.0,"formal_verification":"none","one_line_summary":"DUET improves RLVR by allocating tokens across both prompt selection and rollout length, outperforming full-budget baselines even when using only half the tokens.","context_count":1,"top_context_role":"background","top_context_polarity":"background","context_text":"19274-19286, 2023. 10 [19] Xuefeng Li, Haoyang Zou, and Pengfei Liu. LIMR: Less is more for RL scaling.arXiv preprint arXiv:2502.11886, 2025. [20] Qiang Liu, Lihong Li, Ziyang Tang, and Dengyong Zhou. Breaking the curse of horizon: Infinite-horizon off-policy estimation. InAdvances in Neural Information Processing Systems, 2018. arXiv:1810.12429. [21] Niklas Muennighoff, Zitong Yang, Weijia Shi, Xiang Lisa Li, Li Fei-Fei, Hannaneh Hajishirzi, Luke Zettlemoyer, Percy Liang, Emmanuel Candès, and Tatsunori Hashimoto. s1: Simple test-time scaling. arXiv preprint arXiv:2501.19393, 2025. [22] Jerzy Neyman. On the two different aspects of the representative method: the method of stratified sampling and the method of purposive selection."},{"citing_arxiv_id":"2605.08401","ref_index":47,"ref_count":2,"confidence":0.98,"is_internal_anchor":true,"paper_title":"AIPO: Learning to Reason from Active Interaction","primary_cat":"cs.CL","submitted_at":"2026-05-08T19:06:55+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":6.0,"formal_verification":"none","one_line_summary":"AIPO adds active multi-agent consultation (Verify, Knowledge, Reasoning agents) plus custom importance sampling to RLVR training so LLMs expand their reasoning boundary and then operate without the agents.","context_count":1,"top_context_role":"background","top_context_polarity":"background","context_text":"view of large language model post-training.CoRR, abs/2509.04419, 2025. 1, 2, 5, D.1 13 [46] Lu Ma, Hao Liang, Meiyi Qiang, Lexiang Tang, Xiaochen Ma, Zhen Hao Wong, Junbo Niu, Chengyu Shen, Runming He, Bin Cui, and Wentao Zhang. Learning what reinforcement learning can't: Interleaved online fine-tuning for hardest questions.CoRR, abs/2506.07527, 2025. 2, 5, D.1 [47] Niklas Muennighoff, Zitong Yang, Weijia Shi, Xiang Lisa Li, Li Fei-Fei, Hannaneh Hajishirzi, Luke Zettlemoyer, Percy Liang, Emmanuel J. Candès, and Tatsunori Hashimoto. s1: Simple test-time scaling.CoRR, abs/2501.19393, 2025. 3.1 [48] OpenAI. Learning to reason with llms. https://openai.com/index/learning-to- reason-with-llms/, 2024. Accessed: 2024-09."},{"citing_arxiv_id":"2605.07686","ref_index":9,"ref_count":1,"confidence":0.9,"is_internal_anchor":false,"paper_title":"The Coupling Tax: How Shared Token Budgets Undermine Visible Chain-of-Thought Under Fixed Output Limits","primary_cat":"cs.LG","submitted_at":"2026-05-08T12:54:53+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":8.0,"formal_verification":"none","one_line_summary":"Shared token budgets between visible chain-of-thought and answers create a coupling tax that makes non-thinking competitive on math benchmarks, with a truncation decomposition predicting the crossover and split budgets improving results.","context_count":1,"top_context_role":"baseline","top_context_polarity":"baseline","context_text":"thinking@128 3.6% 144 0.0% 100.0% thinking@256 7.9% 272 0.0% 100.0% thinking@512 18.4% 528 0.7% 99.3% 13 Table 7:Think budget ablation on MATH-500(Qwen3-8B, seed=42). Monotonic improvement at both pilot and full scale. 95% Wilson CIs. Bthink nOverall (%) Stage 1 Stage 2 Stage 3 S3 Acc (%) Nat. Stop Full-scale (n=500) 2048 500 67.2[63.0, 71.2]216 25 259 51.0[44.9, 57.0]25/284 (8.8%) 4096 50074.0[70.0, 77.7]216 127 157 51.6[43.8, 59.3]127/284 (44.7%) ∆ 2048→4096 +6.8∗∗∗-+102- - - ∗∗∗McNemarp=0.0004. Pilot (n=200): 62.5%→73.0%→78.5% atB think∈{1024,2048,4096}. Table 8:Ablation on escalated (hard) samples.Pilot accuracy on 106 escalated MATH- 500 samples from then=200seed-42 run that failed nothink triage. All chains truncated at"},{"citing_arxiv_id":"2605.07588","ref_index":12,"ref_count":1,"confidence":0.9,"is_internal_anchor":false,"paper_title":"Revisiting Transformer Layer Parameterization Through Causal Energy Minimization","primary_cat":"cs.LG","submitted_at":"2026-05-08T11:02:11+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":6.0,"formal_verification":"none","one_line_summary":"CEM recasts Transformer layers as energy minimization steps, enabling constrained parameterizations like weight sharing and low-rank interactions that match standard baselines in 100M-scale language modeling.","context_count":1,"top_context_role":"background","top_context_polarity":"unclear","context_text":"Diagonal-plus-low-rank parameterization. In Section 2.1, we introduced a low-rank parameteri- zation of Ak =W Q⊤ k W K k in the interaction energy, recovering the query and key projections of standard attention. We now ask whether a purely low-rank form is sufficient, and instead propose a diagonal-plus-low-rank parameterization for the matrixA k: Ak = diag(d k) +W Q⊤ k W K k ,(12) where dk ∈R Dh. This augmented parameterization captures key-query interactions that low-rank matrices alone cannot represent, yielding a richer structure for the interaction matrix Ak. The diagonal term enriches the interaction matrix but increases computational cost, so we propose sharing it across heads. A detailed empirical analysis is provided in Figure 3b and more background on this"},{"citing_arxiv_id":"2605.07137","ref_index":17,"ref_count":1,"confidence":0.9,"is_internal_anchor":false,"paper_title":"Adaptive Negative Reinforcement for LLM Reasoning:Dynamically Balancing Correction and Diversity in RLVR","primary_cat":"cs.LG","submitted_at":"2026-05-08T02:13:33+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":4.0,"formal_verification":"none","one_line_summary":"Adaptive scheduling of penalties over training time plus confidence-based weighting of mistakes improves LLM performance on math reasoning benchmarks compared to fixed-penalty negative reinforcement.","context_count":1,"top_context_role":"background","top_context_polarity":"background","context_text":"where α >0 controls sensitivity to confidence and ϵ >0 is a floor ensuring every incorrect sample receives a minimum penalty. Objective:The CW-NSR objective replaces the uniform negative penalty with hardness-weighted penalties: LCW-NSR(θ) =−E x∼D   X y:r(x,y)=1 λ·π θ(y|x)   | {z } λ·LPSR(θ) −E x∼D   X y:r(x,y)=−1 w(y).(−πθ(y|x))   | {z } LC-NSR(θ) . (17) Gradient analysis:The token-level gradient of the CW-NSR loss for an incorrect sample (x,y) with hardnessw(y)with respect to the logitz v of tokenvat stept: For PSR (R= +1): − ∂LPSR ∂zv ∝ \u001aπv(1−π v)ifv=y t (sampled token) −πyt πv ifv̸=y t (unsampled token) (18) For NSR (R=−1): − ∂LC-NSR ∂zv ∝w(y)· \u001a−πv(1−π v)ifv=y t (sampled token) πyt πv ifv̸=y t (unsampled token) (19)"},{"citing_arxiv_id":"2605.06638","ref_index":85,"ref_count":3,"confidence":0.98,"is_internal_anchor":true,"paper_title":"Can RL Teach Long-Horizon Reasoning to LLMs? Expressiveness Is Key","primary_cat":"cs.AI","submitted_at":"2026-05-07T17:48:42+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":6.0,"formal_verification":"none","one_line_summary":"RL training compute for logical reasoning follows a power law with horizon depth whose exponent rises with logical expressiveness, yielding better downstream transfer when models train on richer logics.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2605.06755","ref_index":23,"ref_count":1,"confidence":0.9,"is_internal_anchor":false,"paper_title":"Gradient Extrapolation-Based Policy Optimization","primary_cat":"cs.LG","submitted_at":"2026-05-07T16:20:13+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":6.0,"formal_verification":"none","one_line_summary":"GXPO approximates longer local lookahead in GRPO training via gradient extrapolation from two optimizer steps using three backward passes total, improving pass@1 accuracy by 1.65-5.00 points over GRPO and delivering up to 4x step speedup.","context_count":1,"top_context_role":"background","top_context_polarity":"background","context_text":"We first consider the clean case where the coordinate-wise geometric model is exact. Corollary 2(Diagonal-quadratic GD-surrogate sanity check).Consider the global diagonal quadratic loss L(θ) = 1 2 θ⊤H0θ, H 0 = diag(h1, . . . , hd), h i >0, ηh i ≤1. Assume all nonzero-gradient coordinates are active, finite-precision stabilizers are omitted, and α= 1. Let µ:= min i hi >0, ρ:= (1−ηµ) 2 ∈[0,1). Then one clean GXPO outer step with three backward passes reaches the same point as K+ 1 plain-GD steps: θGXPO new =θ GD K+1, and consequently, afterB∈3Nbackward passes, L \u0010 θGXPO B/3 \u0011 ≤ρ (K+1)B/3 L(θ0), hence, if0< ρ <1, BGXPO =O \u0012 3 K+ 1 log 1 ε \u0013 . This is only an algebraic sanity check: in the easiest case, the extrapolated point lands exactly where multiple GD steps would land. Appendix A.6 proves Corollary 2. Real losses are not diagonal quadratics, so the next result bounds the local error of the GD surrogate. Theorem 3(Local displacement-error bound for the GD surrogate).Suppose K≥2 , L ∈C 3, supξ ∥∇3L(ξ)∥ ≤M 3, and the true GD trajectory satisfies sup 0≤n<K ∥g(θtrue n )∥ ≤G. Letρ ⋆ ≥1andρ ⋆ ≥ ∥I−ηH 0∥. Split coordinates into A={i:|g 0,i|> δ},S=A c. Consider the clean active-set surrogate that uses empirical ratios on A and the observed two-probe displacement on S. If the active empirical ratios and diagonal surrogate rates are bounded by R, then ∥θemp K −θ true K ∥ ≤E off +E ratio +E nonquad, where Eoff comes from off-diagonal Hessian coupling, Eratio from empirical-ratio error and inactive- coordinate fallback, and Enonquad from the Taylor remainder. The explicit constants are given in Theorem 8, Lemma 9, and Corollary 10 in Appendix A.5; together they prove Eoff =O K2η2∥Hoff 0 ∥∥g0∥ρK−2 ⋆ \u0001 , Eratio =O(η 2/δ) +O(η∥g 0,S ∥1) +O η2∥(H0g0)S ∥1 \u0001 , Enonquad =O K3η3M3G2ρK−1 ⋆ \u0001 . This bound gives simple checks for whether GXPO is operating in its intended local regime. The extrapolated displacement should have small error, the error may grow with K but should r"},{"citing_arxiv_id":"2605.05715","ref_index":48,"ref_count":1,"confidence":0.98,"is_internal_anchor":true,"paper_title":"Decodable but Not Corrected by Fixed Residual-Stream Linear Steering: Evidence from Medical LLM Failure Regimes","primary_cat":"cs.AI","submitted_at":"2026-05-07T05:58:38+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":5.0,"formal_verification":"none","one_line_summary":"Overthinking in medical QA is linearly decodable at 71.6% accuracy yet fixed residual-stream steering yields no correction across 29 configurations, while enabling selective abstention with AUROC 0.610.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2605.05561","ref_index":8,"ref_count":1,"confidence":0.98,"is_internal_anchor":true,"paper_title":"BitCal-TTS: Bit-Calibrated Test-Time Scaling for Quantized Reasoning Models","primary_cat":"cs.AI","submitted_at":"2026-05-07T01:10:34+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":5.0,"formal_verification":"none","one_line_summary":"BitCal-TTS raises exact-match accuracy by 3.7 points (7B) and 2.8 points (14B) on small GSM8K shards for 4-bit Qwen2.5 models while cutting premature-stop rates and retaining token savings versus fixed-budget decoding.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2605.03356","ref_index":10,"ref_count":1,"confidence":0.98,"is_internal_anchor":true,"paper_title":"POSTCONDBENCH: Benchmarking Correctness and Completeness in Formal Postcondition Inference","primary_cat":"cs.SE","submitted_at":"2026-05-05T04:29:38+00:00","verdict":"UNVERDICTED","verdict_confidence":"MODERATE","novelty_score":7.0,"formal_verification":"none","one_line_summary":"POSTCONDBENCH is a new multilingual benchmark that evaluates LLM postcondition generation on real code using defect discrimination to assess completeness beyond surface matching.","context_count":1,"top_context_role":"background","top_context_polarity":"background","context_text":"2 languages × 5 models × 3 settings), standalone wins in 51 cases (85%). The gap is larger for completeness: average Comp@1 rises from 0.035 to 0.090 in Python (2.6 ×) and from 0.188 to 0.316 in Java (1.7 ×), suggesting that external dependencies make it harder to produce sufficiently constraining postconditions. Method Length.We bucket methods by lines of code (LoC): [0,20) , [20,40) , and [40,∞) . As shown in Figure 6, performance generally de- 0.2 0.4 0.6Corr@1 C2P N2P F2P [0,20) [20,40) [40, ) #LoC 0.0 0.1 0.2 0.3 0.4Comp@1 [0,20) [20,40) [40, ) #LoC [0,20) [20,40) [40, ) #LoC GPT-5 Claude-4.5 LLaMA-4 Qwen3-32B Gemma-3-27B Figure 6: Corr@1 (top) and Comp@1 (bottom) across target methods grouped by lines of code (LoC). creases with LoC, and the drop is much steeper for completeness. For example, under C2P with Claude-4.5, Comp@1 falls from 0.388 to 0.206 and 0.076 from short to medium and long meth- ods. Aggregated over models, languages, and input settings, the overall average score decreases from 0.296 to 0.205 and then to 0.149 as LoC increases, confirming that longer methods pose a substantially harder generation problem. Method characteristics strongly affect postcon- dition generation: standalone methods are eas- ier than dependency-heavy ones, and longer methods degrade performance monotonically, with completeness being especially sensitive. 5.5 How Stable Is Completeness Under Mutant-Set Changes? We provide ablations to analyze how mutation de- sign affects completeness (details in Appendix F). Operator-level ablation.Each language in- cludes 11 operator-based mutation types. We remove each operator individually and compute Comp@1 (averaged across models). In Python, dropping any single operator changes Comp@1 by at most 0.004; in Java, most operators shift Comp@1 by≤0.001 (see Table 9). Mutation-strength ablation.We further vary mutation \"strength\" by (i) randomly removing 5 of the 11 operators and (ii) randomly removing 50% of LLM-based mutant"},{"citing_arxiv_id":"2605.02035","ref_index":20,"ref_count":1,"confidence":0.9,"is_internal_anchor":false,"paper_title":"VIDA: A dataset for Visually Dependent Ambiguity in Multimodal Machine Translation","primary_cat":"cs.CL","submitted_at":"2026-05-03T19:55:06+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":7.0,"formal_verification":"none","one_line_summary":"VIDA provides 2,500 visually-dependent ambiguous translation examples and span-level disambiguation metrics; CoT-SFT on LVLMs improves out-of-distribution performance over standard SFT.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null}],"limit":50,"offset":0}