Reasoning & Chain-of-Thought - Weekly Roundup

This work proposes SkillSpec, a Hoare-style framework that formulates skill correctness as a specification reasoning problem, and transforms a heterogeneous skill repository into a unified graph representation that aligns descriptions, instructions and code artifacts.

This work proposes MobileVLA-R1 2.0, an RL-enhanced VLA framework that explicitly couples structured embodied reasoning with executable mobile robot control, and introduces a reasoning-conditioned action decoder that maps multimodal reasoning representations to task-level action targets, which are subsequently translated into embodiment-specific commands by robot controllers.

Optimization agents routinely fabricate missing constraints rather than asking users for clarification, but explicit formulation-gap tracking allows agents to reliably recover hidden requirements before writing code.

Chain-of-thought is far more than surface imitation: LLMs carve out discrete, context-dependent geometric subspaces for functional operations like deduction and decomposition, with middle layers encoding the abstract reasoning step rather than the literal tokens.

Long-horizon reasoning traces repeatedly backtrack to early planning steps via predictable query clusters, revealing that a handful of representative "beacon" vectors can guide 5.8× KV cache compression with near-zero quality loss.

This work proposes MobileVLA-R1 2.0, an RL-enhanced VLA framework that explicitly couples structured embodied reasoning with executable mobile robot control, and introduces a reasoning-conditioned action decoder that maps multimodal reasoning representations to task-level action targets, which are subsequently translated into embodiment-specific commands by robot controllers.

This work proposes SkillSpec, a Hoare-style framework that formulates skill correctness as a specification reasoning problem, and transforms a heterogeneous skill repository into a unified graph representation that aligns descriptions, instructions and code artifacts.

Optimization agents routinely fabricate missing constraints rather than asking users for clarification, but explicit formulation-gap tracking allows agents to reliably recover hidden requirements before writing code.

Chain-of-thought is far more than surface imitation: LLMs carve out discrete, context-dependent geometric subspaces for functional operations like deduction and decomposition, with middle layers encoding the abstract reasoning step rather than the literal tokens.

Long-horizon reasoning traces repeatedly backtrack to early planning steps via predictable query clusters, revealing that a handful of representative "beacon" vectors can guide 5.8× KV cache compression with near-zero quality loss.

WearableQA, a benchmark comprising 4,084 10-option multiple-choice questions constructed from the wearable time series, blood biomarkers, and demographics of 200 real users, provides a realistic and diagnostic benchmark for evaluating LLM reasoning over real-world wearable data.

Occamy-1.0, a cost-efficient co-work model obtained by further training the post-trained Qwen3.6-35B-A3B checkpoint, is presented and placed at the low-cost knee of the observed cost--performance Pareto frontier.

Numerical reasoning in HTN planning just got a major upgrade with a new SMT-based encoding that sets a competitive baseline for future advancements.

Only one out of 178 problem-model cells shows prefix-limited reasoning, while most LLMs outperform restarts, highlighting that compute compression is the key driver of performance.

A single scalar from a learned operator can predict convergence in coupled dynamical systems, revealing the limits of traditional backpropagation methods.

Loop Memory Attention enables models to revisit earlier computations, leading to a 2.2% accuracy boost in MathQA tasks compared to fixed loop depths.

Gradient-Aligned Reward transforms how we leverage expert knowledge in reinforcement learning, enabling LLMs to achieve superior reasoning performance with minimal overhead.

Guessing can masquerade as reasoning in GRPO, leading to misleading policy optimization that SIGNBALANCE effectively corrects.

Uncovering causal relationships through symmetries could revolutionize how we approach complex machine-learning tasks beyond IID constraints.

Spurious reasoning persists even after an end-of-think token is injected, complicating the transition to answering in large reasoning models.

Medical image reasoning claims are often inflated by shortcuts, with accuracy plummeting to 5.21% when these cues are mitigated.

Overcoming the 2D-to-4D spatial bottleneck in VLMs does not require native 3D architectures; factorizing visual projections into verifiable planar, depth, and temporal RL objectives delivers immediate 4-6% benchmark gains.

Adaptive routing in programming tutors fails to provide a significant advantage over fixed strategies, challenging assumptions about dynamic response generation in educational AI.

SelfDR achieves high-quality recommendations by distilling reasoning from the same LLM, cutting computational costs without sacrificing performance.

Agents equipped with Semantic Bayesian World Models can make nuanced decisions by treating knowledge as a dynamic web of beliefs rather than static facts.

LLM judges may recognize important reasoning steps, but they still miss critical nuances, highlighting a gap between legibility and true interpretability in AI reasoning.

A novel architecture that decouples topological reasoning from LLMs enables SOC agents to process complex authentication graphs with unprecedented speed and accuracy.

GraFT outperforms fine-tuned models by providing a training-free solution that boosts spatial reasoning in MLLMs by leveraging 3D scene graphs.

Fast and effective counterfactual explanations for routing decisions can be achieved through innovative integer programming techniques, outperforming traditional methods by a significant margin.

ToolDF achieves a remarkable 14.39-point boost in macro-F1 scores over traditional methods while providing interpretable evidence for mixed-authenticity audio deepfake detection.

RuleMem's innovative approach to active memory management leads to a staggering 54.3% improvement in conversational accuracy by transforming dialogue history into reusable logical rules.

LLMs can accurately predict typological features while providing interpretable rationales, even for low-resource languages, when given the right contextual evidence.

Localization-sensitive questions can see up to 40% performance improvement by dynamically selecting reasoning strategies based on question type.

Stateagent transforms video generation by effectively tracking and inferring world states, leading to a 54% improvement in coherence across segments.

Cross-attentive MLLMs already solve the attribute-binding failures that cripple embedding models—and distilling their soft multi-level ranking distributions via Rank-KL unlocks that reasoning for bi-encoders without degrading standard recall.

Today's top video generators score near 0.8 on standard benchmarks like VBench, but none surpass 0.42 when tested on whether two co-occurring objects obey the same basic laws of physics.

Complex scoring algorithms for KV cache compression are largely unnecessary: protecting the initial prompt and dropping reasoning tokens completely at random matches state-of-the-art accuracy with up to 43% higher serving throughput.

Inverting dense self-guidance whenever a verifier rejects a trajectory prevents reasoning models from reinforcing their own hallucinations, enabling stable, energy-based self-improvement without mode collapse.

Propagating local edits across multi-turn artifacts does not require expensive sequential reflection loops—reranking just three parallel samples reliably yields up to a 9.7% consistency boost across open and frontier models.

Cliff reveals that focusing on the first mistake in reasoning can significantly enhance the performance of reinforcement learning models, outperforming traditional methods.

Misleading assertions corrupt LLM medical reasoning more than fabricated evidence, yet are disclosed less often, exposing a critical flaw in model transparency.

Logical validity can be accurately decoded from hidden states of language models, even when their outputs are incorrect, revealing a critical gap between representation and expression.

Continuous mixture collapse in LLMs reveals that even minor distortions can lead to significant reasoning failures, challenging the reliability of latent-state methods.

MLLMs struggle with multi-hop reasoning in document-style images, achieving only 62.83% accuracy compared to over 90% for human annotators.

Structured reasoning in LLMs can dramatically boost diagnostic accuracy in telecom root cause analysis, overcoming the limitations of vanilla models.

SALA achieves superior reasoning alignment by learning flexible, task-specific operations, outperforming traditional methods that rely on rigid matching criteria.

APEx achieves a remarkable 14.7-point improvement over GPT-5.4 by effectively distilling procedural skills for adaptive reasoning in complex question answering.

Placing reasoning traces before context in transformers can boost long-context reasoning accuracy by up to 100% while using exponentially less memory.

A significant reasoning-generation gap exposes the limitations of current generative models, revealing that they often produce outputs that are locally plausible yet globally illogical.

Seed-anchored graph rendering retains all multi-hop evidence in power-grid question answering, achieving a dramatic accuracy boost while adhering to strict context budgets.

LLMs achieve more reliable in-context learning from rules than from examples, challenging the conventional reliance on few-shot prompting.

Nearly orthogonal domain gradients can still wreak havoc in MoE+LoRA setups, but a new approach effectively curbs this interference.

LLMs exhibit a surprising bias towards earlier evidence, which significantly impairs their ability to resolve contextual knowledge conflicts.

Group-level navigation in document retrieval can boost relevance by up to 8% compared to traditional methods.

Transcript-only self-reflection is mathematically incapable of guaranteed improvement without external verification, but grounding reflective updates in environment risk turns multi-agent memory optimization into a provably convergent game.

Traditional causal inference requires hand-crafted pipelines and bespoke estimator training for every dataset, but causal foundation models can now infer treatment effects entirely zero-shot via in-context learning.

Agents can be misled into premature conclusions, but a new method reveals how to alleviate this pressure and improve decision-making.

NeuSOGA transforms complex geometric observations into interpretable symbolic representations, bridging the gap between perception and reasoning in AI.

H2Table achieves a remarkable 22.88% improvement in complex table reasoning, revealing the power of hierarchical hypergraph representations in LLMs.

An anytime solver that achieves 99.90% accuracy on Sudoku-Extreme by leveraging persistent hidden states and maximum corruption during inference upends traditional reasoning models.

SOVER achieves a remarkable 99.33% accuracy in certifying optimization reformulations, revealing that LLMs can enhance formal verification processes significantly.

Dense process supervision can transform RL training by providing step-level rewards that enhance credit assignment, leading to superior performance in complex reasoning tasks.

Fine-tuning a language model into a vision-language model can lead to significant text capability loss, with a novel metric predicting degradation before adaptation.

HypoSearch boosts deep-research agent performance by enabling flexible exploration and evidence comparison, leading to a remarkable 28% increase in accuracy on complex benchmarks.

Iterative refinement of latent states using a small recurrent reasoner can drastically enhance reasoning performance in frozen LLMs, outperforming traditional methods at a fraction of the compute cost.

Tokenization isn’t just a preprocessing step; it’s a hidden driver of model performance and learning dynamics that could redefine how we interpret model evaluations.

Compressed history representations can lead to a staggering 62.7 percentage point drop in reasoning accuracy for clinical LLMs, challenging the assumption that larger models are always better.

Explicit context compilation boosts LLM performance on in-context learning tasks, lifting accuracy from 15.4% to 21.4% on complex benchmarks.

Iterative problem-solving reveals that LLMs, while less accurate, can generate insightful solutions that illuminate failure modes in student coding attempts.

SpatialGuard transforms 3D spatial text-to-image generation from a mere prompt-following exercise into a verifiable planning and validation process, significantly enhancing spatial accuracy.

Change-type awareness in captioning can lead to a dramatic improvement in accuracy, with MEDIC outperforming traditional methods by leveraging specialized reasoning for distinct visual cues.

VerNav slashes decision-stage latency by over 10x while maintaining competitive navigation performance in vision-and-language tasks.

TGR revolutionizes industrial recommendation systems by achieving a staggering 477.8% boost in cold-start new-user engagement through innovative generative techniques.

LLMs struggle to apply programming-language semantics, relying instead on learned priors, especially as program complexity rises.

Leveraging its own failure history, DiagEvo transforms self-play by dynamically guiding question generation to address specific reasoning weaknesses, leading to unprecedented accuracy gains.

Switch Distillation not only enhances reasoning performance by up to 71% but also preserves factual recall, challenging the conventional trade-off in knowledge distillation.

HEART achieves 84% task completion on real-world tasks, outperforming leading models while slashing API costs by up to 85%.

Language models struggle with lifecycle reasoning in complex systems, often faltering as interaction complexity increases, but additional inference effort can unlock significant performance gains.

LRMs mirror human reasoning effort in abductive tasks, revealing shared cognitive challenges and error patterns that could redefine our approach to AI reasoning.

Machine-extracted legal implications can be certified for reliability, but nearly 93% of outputs may still lack sufficient informativeness under real-world error conditions.

A novel neuro-symbolic framework that not only boosts reasoning accuracy in cyber threat intelligence but also provides transparent evidence tracing and reduces hallucinations.

Event-centric memory can boost long-video QA accuracy by over 3 points while slashing inference costs by nearly 64%.

MERGED achieves a 13.79% boost in PR-AUC while being 6x cheaper than larger models, revolutionizing entity resolution in dynamic business environments.

Temporal reasoning benchmarks for Video-LLMs are often mediated by language, leaving room for linguistic shortcuts from option wording, answer correlations, or language priors, so TempCloze is introduced, a video cloze benchmark for evaluating visual temporal reasoning in Video-LLMs.

Tracking both task advantage and actual reward gains can drastically improve the efficiency of multi-task reinforcement learning for LLMs.

As human oversight wanes, LRMs could autonomously evolve, but this shift introduces significant risks like reward hacking and feedback drift.

Continuous reasoning in embedding space can outperform traditional token-based methods while slashing computational costs.

Achieving perfect length generalization with a transformer model of just 280 parameters could revolutionize how we approach algorithmic tasks in neural networks.

S3C-LLM outperforms existing models in molecular analysis by leveraging a self-evolving skill library and a unique two-stage training strategy.

CoMPASS achieves a remarkable balance, enhancing molecular predictions by intelligently integrating LLM insights only when uncertainty is high, leading to improved accuracy without sacrificing reliability.

Even leading MLLMs struggle with object hallucination, revealing critical integration-stage weaknesses that standard benchmarks fail to capture.

Training LLMs with tokens selected by gradient magnitude rather than just entropy boosts reasoning performance across diverse tasks.

Credit-addressable reasoning boosts multimodal geometry accuracy by over 8 points, revealing the critical role of structured learning in complex tasks.

CARVE allows masked diffusion models to dynamically expand their response length, achieving better accuracy and efficiency without retraining.

Many LLMs may possess intact reasoning capabilities hidden beneath output bottlenecks, with simple corrections yielding substantial accuracy gains.

HSRM achieves competitive verification performance with just 2M parameters, challenging the notion that larger models are always necessary for accurate reasoning.

Faithfulness-aware training can slash citation fabrication rates from nearly one-third to just 4.7%, transforming the reliability of AI in clinical settings.

Trajectory geometry can predict reasoning success in LLMs, boosting task performance while cutting costs.

Text-only LLMs can generate images, but their spatial reasoning varies widely, revealing a complex interplay between model architecture and output medium.

AdaPath outperforms traditional methods in biomedical question answering by effectively navigating dense knowledge graphs with minimal surface cues.

Activating task-relevant latent features from high-resource languages can significantly boost reasoning performance in low-resource languages, challenging assumptions about language capabilities.

LCoT-GV reveals that even correct answers can be underpinned by flawed reasoning, and offers a robust method for verifying the integrity of these reasoning chains.

CM2 achieves unprecedented gains in cultural reasoning for MLLMs, revealing the power of a multi-agent approach to interdisciplinary challenges.

Token-level saliency derived from a model's internal reasoning process can drastically improve the efficiency of chain-of-thought compression, outperforming traditional methods.