Topic
reasoning
Think Again or Think Longer? Selective Verification Boosts LLM Accuracy While Cutting Compute Costs
A new preprint on arXiv proposes SEVRA, a serving-layer controller that selectively verifies LLM reasoning outputs. On MATH-500, it achieves 76.3% accuracy — higher than always verifying — while reducing post-generation tokens by 26.8% and harmful flips from 2.2% to 1.0%. The study provides a deployment rule: first tune the initial reasoning budget, then use selective recovery when explicit checks are needed.
Hypergraph Reasoning Framework Boosts Semantic Communication Accuracy by 36.6%
A new hypergraph-based framework, HISR, improves implicit semantic interpretation accuracy by up to 36.6% over existing methods by capturing higher-order relationships among entities, enabling robust performance even under noisy channel conditions.
Reinforcement-Aware Knowledge Distillation Boosts LLM Reasoning Efficiency
Researchers propose RL-aware distillation (RLAD) to address distribution mismatch and objective interference in knowledge distillation for LLM reasoning. The method uses Trust Region Ratio Distillation (TRRD) to selectively imitate teacher policies during reinforcement learning. RLAD outperforms offline distillation, standard GRPO, and KL-based on-policy distillation across logic and math benchmarks.
Beyond Reasoning Gains: Mitigating General-Capability Forgetting in Large Reasoning Models
A new research paper from arXiv shows that reinforcement learning with verifiable rewards (RLVR) can cause large reasoning models to forget foundational capabilities like perception and faithfulness. The authors propose RECAP, a replay strategy with dynamic objective reweighting that preserves general knowledge while maintaining reasoning gains.
FM-Agent: New Framework Automates Formal Code Verification for Large-Scale LLM-Generated Software
FM-Agent, a new framework from researchers, automates compositional reasoning for large-scale systems using LLMs. It generates function-level specifications from caller expectations, enabling verification against natural-language intent. In evaluation, it found 522 new bugs in systems up to 143,000 lines of code within 2 days.
Independent Combinatorial Tokens Framework Boosts LLM Reasoning Performance by Up to 14.9%
Researchers propose the Independent Combinatorial Tokens (ICT) framework to resolve entropy collapse and explosion in LLM reasoning. By focusing on token-level distributional deviations using Jensen-Shannon divergence, ICT achieves average pass@4 improvement of 4.58% and up to 14.9% over baselines.
New Method LUCID Detects Hallucinations in LLM-Based Knowledge Graph Reasoning
Researchers introduce LUCID, the first hallucination detection method designed for large language model-based knowledge graph reasoning. By jointly leveraging attention scores, KG semantics, and structural information via a graph neural network, LUCID achieves state-of-the-art performance across nine datasets compared to 15 baselines. The method addresses a critical gap where existing detection techniques overlook structural information in knowledge graphs.
QMFOL Benchmark Reveals LLM Reasoning Degrades with Logical Complexity, New Framework Enables Precise Evaluation
A new automated framework called QMFOL generates deductive reasoning tasks with quantifiable logical complexity, enabling precise evaluation of LLM reasoning. The associated benchmark, QMFOLBench, comprises 2,880 instances across 960 configurations. Evaluations on six large reasoning models (LRMs) and two LLMs show performance degrades and computational overhead increases with rising logical complexity, with models performing better on True-labeled tasks than False or Unknown ones.
MedAI Study Evaluates TxAgent's Therapeutic Reasoning in NeurIPS CURE-Bench Competition
A MedAI study evaluated TxAgent, an agentic AI system for therapeutic reasoning, in the NeurIPS CURE-Bench 2025 Challenge. TxAgent uses a fine-tuned Llama-3.1-8B model with iterative retrieval-augmented generation and a unified biomedical tool suite. The work was awarded the Excellence Award in Open Science.
Research Shows Code Execution Outperforms Natural Language for AI Algorithmic Reasoning
A new research paper from arXiv investigates whether code or natural language is more effective for tool-augmented language models performing algorithmic reasoning. By separating intermediate representation from execution mechanism, the study finds that deterministic code execution outperforms natural-language reasoning by 31.6 percentage points, while changing the intermediate representation alone yields only a 0.15pp difference. Results suggest performance gains require reliable external execution.
SorryDB Benchmark Tests AI Provers on Real-World Lean Theorem Completion Tasks
Researchers present SorryDB, a benchmark of open Lean tasks from 78 GitHub projects. Evaluating a snapshot of 1000 tasks, they show current approaches are complementary, with Gemini Flash-based agentic methods leading but not outperforming all others.
Multi-Sequence Verifiers Cut Inference Latency in Half for LLM Reasoning
A new paper by Kim et al. introduces the Multi-Sequence Verifier (MSV), a lightweight verifier that improves calibration for parallel test-time scaling in large language models. MSV enhances best-of-N selection accuracy by up to 6% and enables early-stopping strategies that achieve the same accuracy with less than half the inference latency.
New Framework TRACED Evaluates LLM Reasoning Using Geometric Stability and Progress
A new research framework called TRACED evaluates LLM reasoning quality by analyzing geometric progress and stability of reasoning traces. It distinguishes correct reasoning from hallucinations based on trajectory patterns, offering a more robust evaluation method than scalar probabilities.
AgenticRec: A Recommender Framework That Aligns LLM Reasoning with User Preferences
Researchers propose AgenticRec, a framework that treats recommendation as a tool-integrated reasoning process. It employs a two-stage training paradigm to overcome misalignment between LLM reasoning trajectories and recommendation feedback, improving fine-grained preference distinction.
New Research Shows Chain-of-Thought Reasoning Should Be Selective, Not Default, for LLMs
A research paper on arXiv argues that chain-of-thought (CoT) reasoning should not be the default for large language models. The authors propose EDRM, a training-free routing framework that uses early decoding entropy to decide when to use CoT, achieving up to 55% token reduction and accuracy improvements across 15 benchmarks.
Study: LLM Accuracy Declines Predictably as Reasoning Steps Increase in Clinical AI Tasks
A study on arXiv introduces a hop-count taxonomy to predict LLM failure on clinical question answering. Tests across Claude and GPT models show monotone accuracy decline with reasoning depth, with extended thinking failing to flatten the curve.
Less is More: Improving LLM Reasoning with Minimal Test-Time Intervention
Researchers propose Minimal Test-Time Intervention (MTI), a training-free method that enhances large language model reasoning by focusing on localized, high-entropy tokens. MTI achieves +9.28% average improvement on six benchmarks for DeepSeek-R1-7B and +11.25% on AIME2024 for Ling-mini-2.0, with minimal computational cost.
Tyler Framework Boosts LLM Reasoning by Up to 14 Points with Smarter Compute Allocation
A new framework called Tyler introduces typed latent reasoning for large language models, learning when to invoke latent computation and how much to allocate. On three backbone LLMs, Tyler improved accuracy by up to 14.49 points over chain-of-thought prompting and up to 4.30 points over competing baselines, while reducing forgetting.
Multi-Agent Peer-Reviewed Reasoning Boosts LLM Accuracy in Medical Question Answering
Researchers designed a multi-agent peer-reviewed reasoning method for medical question answering, where multiple LLMs generate and evaluate each other's chain-of-thought reasoning. Experiments with five models on three benchmarks showed the approach consistently outperforms single-model reasoning and majority voting, achieving best accuracy of 0.820. The method scales effectively and improves interpretability.
XMedFusion: A Knowledge-Guided Multimodal Perception and Reasoning Framework for Autonomous Medical Systems
Researchers introduce XMedFusion, a knowledge-guided multimodal perception and reasoning framework for autonomous medical systems. The framework decomposes visual information into coordinated agents, achieving significant improvements in radiology report generation metrics on a public chest radiograph dataset.
DeepRoot Multi-Agent System Enables Therapeutic Reasoning Over Historical Medical Texts with 47.6% Accuracy
DeepRoot is a multi-agent LLM system that jointly builds and utilizes a verified knowledge graph for therapeutic reasoning over historical medical texts. Applied to the Shen Nong Ben Cao Jing, it recovers 10 of 21 held-out compound-disease treatment pairs at R@20 (47.6%), significantly outperforming a raw corpus LLM (4.8%) and random baseline (2.4%). The system also reduces hallucination to 7-10% compared to 87% for tool-using LLMs, offering a scalable method for mining historical medical knowledge.
New Hindsight Self-Distillation Method Improves LLM Reasoning by Localizing Credit at Divergence Points
A new method called Hindsight Self-Distillation (HSD) improves large language model reasoning by conditioning the teacher on a successful peer rollout. This localizes the credit signal at the divergence point between failed and successful rollouts, leading to state-of-the-art results on math and code benchmarks with Qwen3-8B and Qwen3-32B models.
New Defense Keeps Attack Success Rate Below 4% for Adaptive Prompt Injection on LLM Agents
Researchers propose RETA, a training-based defense that grounds LLM agent security on user tasks rather than attack patterns. Using chain-of-thought reasoning and red-teaming with diversity reward, RETA keeps average attack success rate below 4% across six adaptive attacks while preserving utility.
AdaSTORM Breakthrough Scales LLM Reasoning to Thousand-Node Dynamic Graphs, Paves Way for Supply Chain AI
AdaSTORM, a new multi-agent AI framework, scales large language model reasoning to dynamic graphs of up to thousand nodes with over 90% accuracy. The approach uses adaptive partitioning and collaborative reasoning to overcome limitations of current LLMs, which can only handle tens of nodes. This breakthrough could enable AI-driven analysis of complex, evolving networks such as supply chains.
VibeThinker-3B: Small Language Model Matches Giants in Verifiable Reasoning, According to arXiv Paper
A new technical report on arXiv introduces VibeThinker-3B, a compact 3B-parameter language model that achieves verifiable reasoning scores comparable to models orders of magnitude larger, including DeepSeek V3.2, GLM-5, and Gemini 3 Pro. The model uses a Spectrum-to-Signal post-training paradigm and achieves 94.3 on AIME26 and 80.2% Pass@1 on LiveCodeBench v6.
Limited Marginal Benefit of Reasoning-Heavy LLMs in ESG Scoring: Study on Japanese Firms
A 4-model consensus study on 10 Japanese listed firms found that reasoning-heavy LLMs add little value over cheaper alternatives in ESG narrative scoring, with a mean absolute deviation of only 0.38 on a 5-point scale and 5.6x higher cost.
AdaMame: New Training Recipe Solves Language Collapse in Multilingual Reasoning Models
AdaMame, a two-stage training recipe for multilingual mathematical reasoning, addresses language collapse in large reasoning models. It adaptively aligns reasoning language to the query language without compromising accuracy, achieving Pareto-optimal performance across 12 languages.
New Self-Enhanced Fine-Tuning Method Boosts Text-to-SQL Reasoning and Generalization
Researchers propose CoTE-SQL, a self-enhanced fine-tuning method that improves text-to-SQL generation by integrating reasoning traces, structured chain-of-thought prompting, and execution error correction. The approach achieves state-of-the-art results on Bird and Spider benchmarks, particularly on complex queries.
New Benchmark IRTS-ToolBench Tests LLMs on Irregular Time Series Question Answering
A research paper introduces IRTS-ToolBench, a benchmark of 1,700 questions spanning 10 task types across 13 domains to evaluate large language models (LLMs) and AI agents on irregular time series question answering (TSQA). The benchmark addresses a gap in existing TSQA benchmarks that assume regular sampling, providing standardized inputs and a reproducible evaluation protocol for verifiable agentic data science.
Latent Thought Flow: Efficient Reasoning in LLMs Cuts Cost and Boosts Accuracy
Researchers propose Latent Thought Flow (LTF), a method that models LLM reasoning as continuous trajectories in latent space, using GFlowNet and entropy-weighted objectives. LTF outperforms explicit Chain-of-Thought and latent reasoning baselines, achieving 9.5% higher accuracy while cutting reasoning length by 27.2%, addressing the linguistic bottleneck that inflates inference costs.
Think-at-Hard: Selective Latent Iterations Boost LLM Reasoning Accuracy by Up to 6.8%
A new research paper proposes Think-at-Hard (TaH), a looped transformer that selectively performs latent iterations only on tokens likely to be incorrect. By skipping iterations on 93% of tokens, TaH outperforms always-iterate models by 3.8-4.4% and single-iteration baselines by up to 6.8%, while requiring negligible extra parameters.
CycliST Benchmark Reveals Video Language Models Struggle with Cyclical State Transitions
The CycliST benchmark, introduced by a team of researchers, evaluates Video Language Models on cyclical state transitions. Results show current VLMs struggle to detect and reason about periodic patterns, with no single model performing consistently across all tasks.
PrologMCP: A Standardized Prolog Tool Interface That Boosts LLM Agents’ Deductive Accuracy
A team of researchers introduced PrologMCP, an open-source server that exposes Prolog as a stateful tool through the Model Context Protocol, allowing LLM agents to delegate deductive reasoning tasks. In evaluations on the PARARULE-Plus benchmark, an agent powered by PrologMCP achieved accuracy of 1.00 on a general sample, matching or exceeding reasoning LLMs, and 1.00/0.99 on a challenging subset where reasoning models dropped to 0.95/0.94.
The Quality-Utility Paradox: Why High-Reward Data Impairs Small Model Mathematical Reasoning
A research paper identifies a 'Quality-Utility Paradox' in mathematical reasoning distillation: data refined by stronger models (Oracle) receives high reward scores but impairs small model performance compared to using the model's own self-generated traces. The authors propose Style-Aligned Refinement to preserve native reasoning patterns while incorporating logical corrections.
Semi-Supervised Framework Scales LLM Reasoning Using 10-15x Fewer Labels Than Traditional Methods
A new semi-supervised framework for training LLM reasoning uses a lightweight verifier to judge reasoning quality, requiring only a few labeled samples. Experiments on math problems and visual question answering show accuracy comparable to 10-15x more labeled data. The method could reduce the cost of building large-scale reasoning datasets.