Reasoning & Chain-of-Thought - Weekly Roundup

LLMs leverage a difference-making logic akin to experimental methods, reshaping our understanding of how they infer causal relationships from text.

ACOER reduces token generation by over 60% while boosting accuracy, solving the reward collapse problem that plagues traditional efficiency training methods.

Targeting only the gaps in information, GDP-RAG achieves unprecedented accuracy in multi-hop question answering while slashing computational costs.

Off-policy degree in RLVR updates can drastically change which tokens drive learning, leading to a new adaptive method that outperforms traditional baselines.

VADAOrchestra achieves verifiable and adaptable decision-making workflows by seamlessly integrating LLMs with symbolic reasoning, outperforming traditional systems in real-world applications.

LLMs leverage a difference-making logic akin to experimental methods, reshaping our understanding of how they infer causal relationships from text.

ACOER reduces token generation by over 60% while boosting accuracy, solving the reward collapse problem that plagues traditional efficiency training methods.

Targeting only the gaps in information, GDP-RAG achieves unprecedented accuracy in multi-hop question answering while slashing computational costs.

Off-policy degree in RLVR updates can drastically change which tokens drive learning, leading to a new adaptive method that outperforms traditional baselines.

VADAOrchestra achieves verifiable and adaptable decision-making workflows by seamlessly integrating LLMs with symbolic reasoning, outperforming traditional systems in real-world applications.

Retrieval-based learning strategies significantly outperform learner-directed study methods in retaining knowledge gained from AI-assisted pretesting.

Contextual tunneling in LLMs can be overcome, leading to more reliable and physically grounded materials discovery through the innovative ARIA framework.

CoT reasoning boosts verbal reasoning but falters in visual tasks, revealing a critical gap in multimodal AI capabilities.

Bypassing final-layer perturbations can significantly enhance reasoning capabilities in aligned LLMs, achieving better performance with zero memory overhead.

When faced with search-dependent reasoning tasks, models struggle to learn effective chain-of-thought strategies, highlighting a critical limitation in current training paradigms.

CalVerT boosts QA performance by equipping agents with calibrated self-confidence and grounding scores, reducing both erroneous confident answers and unnecessary information retrieval.

ASYS reveals a novel way to automate the discovery of analytical forms for PDEs, producing interpretable solutions where none existed before.

Counterfactual reasoning in neural probabilistic logic just got a major upgrade, achieving 2.14× faster inference while tackling biases in intervention estimates.

Two models can achieve the same accuracy but differ dramatically in logical compliance, revealing a critical oversight in standard evaluation practices.

RACL achieves up to 8.337% cost savings in vehicle routing by intelligently guiding metaheuristic optimizers without altering existing constraints.

Hypergraphs can boost semantic interpretation accuracy by over 36% in communication systems, tackling the pitfalls of traditional pairwise graph representations.

Tactic-level supervision from symbolic proof assistants can significantly outperform traditional outcome-only reinforcement learning approaches in theorem proving tasks.

LLMs exhibit significant brittleness in combinatorial reasoning, particularly with ordered and indistinguishable elements, highlighting critical gaps in their understanding of constraints.

CoT transformers can simulate Word RAM algorithms with poly-logarithmic overhead, revolutionizing our understanding of their computational efficiency.

CARE transforms the approach to reasoning length in video-MLLMs, enabling models to adaptively balance exploration and efficiency based on their evolving competence.

Performance of large reasoning models drops significantly as logical complexity rises, revealing critical gaps in current evaluation benchmarks.

TimeProVe achieves a remarkable 7.3% performance boost on long video reasoning tasks while slashing inference costs by 93%.

Spatial reasoning can be transformed from isolated frame predictions to dynamic scene understanding, significantly boosting performance in multi-view and video tasks.

MedRLM transforms clinical decision support by enabling recursive reasoning over complex patient data, significantly enhancing diagnostic accuracy and referral efficiency.

Rigid geometric compression can collapse reasoning space, while generative reconstruction preserves semantic integrity and enhances reasoning accuracy.

VIMPO achieves superior performance in reasoning tasks by enabling fine-grained credit assignment without the complexities of a critic, redefining the landscape of reinforcement learning for LLMs.

ADaPT enables a single model to flexibly navigate the efficiency-performance trade-off, achieving significant cost savings without sacrificing reasoning quality.

Effect quantales reveal a surprising new dimension to abstract interpretations, shifting the focus from states to event occurrences.

Selective verification can boost accuracy while slashing verification costs, but sometimes a longer initial solve is the more efficient choice.

A single generalist model outperforms specialized systems, achieving over 35% improvement in real-world robotic task success.

CogniRoute outperforms existing models by over 15 percentage points in social video QA, revealing the critical role of cognitive schema in multimodal reasoning.

Writing user-specific facts as local edits results in a 33,000x smaller memory footprint while enhancing reasoning accuracy by 5.6x compared to traditional methods.

FAPO achieves a remarkable +33.8 percentage point gain in performance by seamlessly transitioning from prompt optimization to structural adjustments in LLM pipelines.

OmniAgent not only outperforms larger models but also scales performance with reasoning turns, revolutionizing how we approach video understanding.

Diffusion-Proof not only surpasses AR LLMs in theorem proving but also solves challenging problems that state-of-the-art models fail to address.

MAST achieves targeted forgetting in RLVR-induced reasoning with minimal impact on performance, preserving critical task accuracy while effectively unlearning unwanted knowledge.

Iterative resampling with LLM conditionals can outperform traditional autoregressive generation in structured probabilistic inference.

A noisy predictor that barely beats random guessing can dramatically enhance the efficiency of exact exponential algorithms for NP-hard problems.

Self-conditioning on verified trajectories boosts reinforcement learning performance by over 8%, revealing the power of internal feedback in credit assignment.

Combining multiple properties in ABox abduction doesn't increase complexity, opening new avenues for efficient knowledge base explanations.

VLMs falter in strategic reasoning, with performance dropping sharply in complex RTS scenarios that require tight coordination and long-horizon planning.

A simple data recipe can outperform complex reward engineering in enhancing long-context reasoning for large language models.

GraphPO slashes redundancy in reasoning model training, enabling more efficient exploration and improved performance on complex tasks.

Counterfactual reasoning is the key to unlocking a 5.5% accuracy boost in pragmatic language understanding for LLMs, challenging their traditional reliance on literal interpretations.

Visual-OPSD achieves a remarkable 14.3x speedup while boosting accuracy by over 3 percentage points, revealing the untapped potential of reasoning in visual thought generation.

Achieving nearly 100% annotation acceptance on flowcharts and a 35.1% accuracy boost for VLMs, ScreenAnnotator redefines data annotation for complex visual reasoning tasks.

MAFP reveals that treating stakeholder stances as agents in a game-theoretic framework can drastically improve decision quality in complex scenarios.

Social reasoning in language models is rooted in distinct training data, with targeted unlearning revealing its vulnerability to data removal.

Training with large block sizes cripples reasoning performance, but a novel curriculum approach unlocks strong reasoning capabilities in diffusion models.

Rubric-Conditioned Self-Distillation transforms vague scalar rewards into precise, actionable feedback, leading to significant improvements in reasoning model performance.

Adaptive weighting in model merging can drastically improve multilingual reasoning performance, outperforming traditional methods across 21 languages.

Spectral diagnostics reveal that hallucinations in LLMs can be detected with unprecedented accuracy by analyzing the thermodynamic properties of attention graphs.

The interaction between language models and formal solvers can compromise the soundness of answers, even when using certificate gating to ensure correctness.

JustDiag reveals that accountability in root cause analysis hinges on explicit justification rather than just fluent conclusions.

VERITAS reveals that incorporating verifier feedback can boost theorem proving success rates by over 10%, challenging the efficacy of traditional binary pass/fail approaches.

Achieving nearly identical accuracy to full-VLM systems while slashing invocation rates and costs by nearly 30% could redefine efficiency in multimodal AI applications.

ViGOS reveals that decoupling perception from reasoning can significantly enhance multimodal model robustness against shortcut learning.

By focusing on correcting "near-miss" answers, REVES achieves a remarkable +6.5 point improvement over standard RL methods, showcasing a new way to enhance LLM reasoning without extensive computational costs.

Self-distillation can be transformed from mere imitation of a privileged distribution to a powerful tool for diagnosing and correcting specific reasoning failures in large language models.

Large language models struggle with logical reasoning in Chinese, revealing a persistent performance gap compared to English that challenges their multilingual capabilities.

Training on compound reasoning traces yields better generalization than isolated atomic modules, reshaping our understanding of how LLMs can learn to reason.

CERS leverages Chain-of-Thought reasoning to enhance medical image segmentation, significantly improving accuracy in clinically challenging scenarios where visual cues alone fall short.

Transforming fleeting reasoning traces into durable knowledge can elevate LLM performance beyond traditional training paradigms.

Early belief drift in LLMs can be corrected through innovative resampling techniques, leading to a more stable and coherent predictive process.

ReLAR achieves improved reasoning stability and accuracy in LLMs with significantly lower inference costs, challenging the notion that complex reasoning always requires elaborate chain-of-thought processes.

Fixed-point convergence enables adaptive computation in reasoning tasks, allowing models to efficiently tackle complex challenges without unnecessary resource expenditure.

Current AI systems struggle with research-level mathematics, often failing to match human problem-solving capabilities on complex tasks.

Enhanced IsabeLLM can now verify Bitcoin's Proof of Work consensus more effectively, showcasing AI's transformative role in formal verification.

Domain-specific tools can deliver 90% accuracy in optical network management while slashing token usage by three times compared to their generic counterparts.

Small initialization can dramatically enhance reasoning performance in large language models, revealing a new lever for improving AI capabilities.

A neuro-symbolic framework that achieves 92% accuracy in strategy synthesis by combining LLMs with formal verification methods could revolutionize how agents strategize in complex multi-agent environments.

Structural properties of GNNs reveal deep connections to graded modal logic, offering a new lens to understand their expressiveness.

Fine-grained visual dependencies can drastically improve multimodal reasoning accuracy in mathematical problem-solving, challenging the notion that visual inputs are merely auxiliary.

A single model leveraging homotopy type theory can outperform diversity-trained ensembles in reasoning tasks, revealing the power of symmetry-aware inference.

FlowRAG transforms how we approach multi-hop reasoning by leveraging a quad-level graph structure that enhances both semantic recall and explicit reasoning paths.

Routing SQL queries based on complexity allows DecoSearch to achieve unprecedented execution accuracy while using an order of magnitude fewer tokens than traditional methods.

E^3RL not only overcomes the autoregressive curse but also enhances LLMs' reasoning capabilities, achieving up to 6.514% better performance than previous state-of-the-art models.

Reducing reasoning tokens while boosting accuracy, SuCo transforms how LRMs approach problem-solving by focusing on sufficiency rather than excess.

Task decomposition quality is the critical bottleneck in skill composition, and our Iterative Skill-Aware Decomposition method dramatically boosts accuracy and retrieval performance.

Overthinking in RL models can amplify during training, but Dynamic Rollout Editing effectively curtails this behavior, enhancing reasoning efficiency.

Even top-performing Vision Language Models miss 30% of logical anomalies, revealing critical gaps in their reasoning abilities.

CoTIR achieves superior image restoration by internalizing reasoning processes, outperforming traditional methods even in complex degradation settings.

A new benchmark and model reveal that even minimal cross-scale supervision can dramatically enhance pathological image interpretation.

ThinkingVLA achieves a breakthrough in robotic manipulation by seamlessly interleaving visual and textual reasoning, leading to superior performance in long-horizon tasks.

LALMs can boost their temporal reasoning accuracy by 3.2% simply by better redistributing attention across audio tokens rather than relying on textual cues.

Coordinating neural planning with symbolic execution, Quarry boosts automated proof success rates by up to 13% while keeping costs predictable.

Despite achieving an 83.4% accuracy, the best AI model still falters in critical reasoning areas, exposing the limitations of current embodied AI capabilities.

SR-REAL's dual-path reasoning framework allows spatial VLMs to excel in both linguistic deduction and 3D geometric inference, significantly enhancing performance on complex spatial reasoning tasks.

Without a benchmark for doctrinal legal reasoning, compliance with the EU AI Act's accuracy requirements remains unachievable, risking the integrity of AI in the judicial system.

Multi-agent response planning can elevate intrusion response effectiveness while ensuring analyst oversight and safety.

LLMs can seamlessly generate formal proofs with a new system that mimics natural mathematical language, bridging the gap between human intuition and formal verification.

DeSRPA achieves superior emotional consistency and speech naturalness in SRPAs without the need for extensive retraining, challenging the dominance of end-to-end fine-tuning approaches.

VideoCFR not only boosts performance in video reasoning tasks but also reveals the critical visual evidence driving model decisions without relying on human annotations.

English LLMs dominate math reasoning with a richer set of parameters, while lower-resource languages struggle with significant gaps.

Layer span variation can unlock a new dimension of deterministic rollout diversity, boosting performance by over 10 percentage points on reasoning tasks.

A hybrid architecture that leverages deliberative planning can significantly outperform traditional reactive RL policies in complex environments.

GRACE reveals that existing models struggle with reasoning fidelity, exposing a critical gap in how we evaluate AI's contextual understanding.

Explicit evidence graphs in VeriGraph enable LLMs to achieve 87.61% claim grounding, transforming how we verify AI-generated conclusions.

Compact models like VibeThinker-3B can achieve frontier-level reasoning performance, rivaling much larger counterparts without losing controllability.

Chain-of-thought reasoning amplifies anti-Muslim bias in LLMs, revealing critical gaps in existing mitigation strategies across realistic deployment scenarios.