arXiv AI
Jul 14
Filtered Reasoning Score: Evaluating Reasoning Quality on a Model's Most-Confident Traces
arXiv:2604. 11996v2 Announce Type: replace-cross Abstract: Should we trust Large Language Models (LLMs) with high accuracy?
ReFIne: A Framework for Trustworthy Large Reasoning Models with Reliability, Faithfulness, and Interpretability
ReFIne is a training framework that augments large reasoning models with three trustworthiness properties: interpretability, faithfulness, and reliability. It combines supervised fine‑tuning with GRPO to produce structured, tag‑based reasoning traces, explicitly disclose decisive information, and provide self‑assessments of soundness and confidence. Applied to Qwen3 models, ReFIne improves interpretability by 44.0 %, faithfulness by 18.8 %, and reliability by 42.4 % on mathematical benchmarks.
Related stories
arXiv AI
Jun 3
Quantifying Faithful Confidence Expression in Large Reasoning Models
arXiv:2606. 03969v1 Announce Type: cross Abstract: Reliable uncertainty communication is critical to the trustworthiness of LLMs, yet faithful calibration (FC)--the alignment between models' intrinsic and (linguistically) expressed confidence--is a persistent failure mode.
arXiv Machine Learning
Aug 11
Are Latent Reasoning Models Easily Interpretable?
arXiv:2604. 04902v2 Announce Type: replace Abstract: Latent reasoning models (LRMs) have attracted significant research interest due to their low inference cost (relative to explicit reasoning models) and theoretical ability to explore multiple reasoning paths in parallel.
arXiv AI
Jun 11
Calibration Drift Under Reasoning: How Chain-of-Thought Budgets Induce Overconfidence in Large Language Models
arXiv:2606. 11211v1 Announce Type: cross Abstract: The ability of large language models (LLMs) to express calibrated uncertainty is important for safe deployment.
arXiv Machine Learning
Sep 4
Legibility is Not Interpretability: Comparing Judged and Actual Importance in Chain-Of-Thought Reasoning
The paper investigates whether the text of chain‑of‑thought reasoning steps actually reflects their true importance for a model’s final answer. By defining step importance as the advantage in expected reward when a step is included, the authors use Monte Carlo rollouts to estimate ground truth and then test whether large language model judges can identify high‑advantage steps. They find that capable LLMs can beat a prevalence baseline but still fall far short of a noise ceiling, and that fine‑tuning a step‑level critic improves detection for incorrect responses but remains distant from the ceiling for correct ones, indicating that step importance is only partially recoverable from the reasoning trace text.
arXiv AI
Oct 1
Making LLMs Say What They Think: Measuring and Improving CoT-Interpretability Alignment
The paper introduces CoT-Interpretability Alignment (CIA), a metric that quantifies how well a large language model’s chain-of-thought (CoT) explanations match its internal reasoning processes. Evaluated on two-hop question answering, hint intervention, and integer multiplication across three LLMs, the study finds limited alignment (44.8–75.9%) and demonstrates that post‑training with a reward combining task accuracy and parametric faithfulness can substantially improve CoT faithfulness without sacrificing accuracy. The authors provide a framework for auditing CoT faithfulness and a pathway to making explicit reasoning more trustworthy, with code and data publicly available.
arXiv AI
Jun 9
Correct Is Not Enough: Training Reasoning Planners with Executor-Grounded Rewards
arXiv:2605. 03862v4 Announce Type: replace Abstract: Reinforcement learning with verifiable rewards has become a common way to improve explicit reasoning in large language models, but final-answer correctness alone does not reveal whether the reasoning trace is faithful, reliable, or useful to the model that consumes it.
arXiv Machine Learning
Sep 10
Do Reasoning Representations Help Humans Evaluate LLM Outputs?
The paper investigates whether reasoning representations—explanations for large language model outputs—aid humans in evaluating those outputs. A controlled human study tested six reasoning formats across tasks of varying complexity, measuring structural understanding, error detection, and trust calibration. Results revealed a mismatch: participants favored planning- and decomposition-based representations, yet simpler chain-of-thought traces better supported verification, trust, and interpretability, while preferred formats increased calibration risks.
arXiv AI
Oct 12
ReasoningShield: Safety Detection over Reasoning Traces of Large Reasoning Models
ReasoningShield introduces a benchmark and lightweight models for detecting unsafe content in the intermediate reasoning steps of Large Reasoning Models (LRMs). The benchmark contains 9.2K annotated (Query, CoT) pairs across 10 risk categories, 3 safety levels, and 8 LRMs, and the proposed models achieve 91.8% F1, outperforming existing moderation tools by significant margins. The study highlights that current moderation tools degrade by up to 35.2% on Chain-of-Thought reasoning compared to traditional answer moderation.
arXiv AI
Aug 17
The Metacognitive Bottleneck: Japanese Riddles Reveal Fundamental Limits of Machine Insight and Self-Evaluation in Reasoning AI
arXiv:2509. 14704v3 Announce Type: replace Abstract: Benchmark saturation and training-data contamination increasingly obscure whether reported gains in large language models (LLMs) reflect genuine advances in reasoning or familiarity with recurring patterns in benchmark problems.
arXiv Machine Learning
Oct 2
Less Data Approximates More: Earning Faithful Confidence in High-Stakes Domains
arXiv:2604.08454v2 Announce Type: replace Abstract: Large language models are increasingly deployed in high-stakes domains, where confident yet incorrect inferences may cause severe real-world harm,...
arXiv AI
Aug 12
TrAC: Trace-Conditioned Answer Consistency for Efficient Uncertainty Quantification in LLMs
arXiv:2608. 00422v2 Announce Type: replace Abstract: Large language models (LLMs) can generate fluent reasoning traces that nevertheless lead to incorrect answers, making response-level uncertainty estimation important for abstention, human review, and adaptive compute allocation.