Reasoning & Chain-of-Thought - Weekly Roundup

Autonomous research agents can now produce publication-grade manuscripts in frontier physics, achieving substantive findings while navigating complex literature landscapes.

AI legal advice is perceived as both more objective and less comprehensive, revealing a nuanced public response that challenges the notion of algorithm aversion.

LLMs possess an internal estimate of their remaining output length, revealing a surprising layer of planning in their generation process.

Privileged self-distillation can paradoxically hinder thinking models, leading to a 17% drop in accuracy on long reasoning tasks due to its impact on learning dynamics.

Verbalization, not knowledge, drives fabrication in LLMs, revealing a critical leverage point for improving mathematical reasoning accuracy.

AI legal advice is perceived as both more objective and less comprehensive, revealing a nuanced public response that challenges the notion of algorithm aversion.

LLMs possess an internal estimate of their remaining output length, revealing a surprising layer of planning in their generation process.

Privileged self-distillation can paradoxically hinder thinking models, leading to a 17% drop in accuracy on long reasoning tasks due to its impact on learning dynamics.

Verbalization, not knowledge, drives fabrication in LLMs, revealing a critical leverage point for improving mathematical reasoning accuracy.

TacReasoner outperforms larger models in tactile reasoning tasks, achieving competitive results with fewer parameters.

AssemCAD transforms the landscape of CAD assembly generation by ensuring that mechanical assemblies are not only generated but also validated against engineering principles, achieving unprecedented levels of physical consistency.

Targeted feedback can slash calculation errors in small language models from 56.9% to 23.5%, revolutionizing their physics reasoning abilities.

ClassicLogic reveals that agents can be systematically evaluated on their ability to compose complex problem-solving strategies, a crucial aspect often overlooked in existing benchmarks.

As AI systems become more capable and opaque, the role of symbolic methods shifts from computation to essential interfaces for human oversight and trust.

LLMs can now be trained to prioritize task constraints intrinsically, resulting in a dramatic improvement in planning reliability.

EventCoT achieves state-of-the-art RTL performance with fewer visual tokens, revolutionizing how we approach temporal reasoning in videos.

Adversarial documents can not only mislead deep research agents but also shift poisoned content from overt framing into seemingly factual premises, complicating detection.

Faithful reasoning in VLA models can boost policy responsiveness to rare scenarios by 1.6x compared to state-of-the-art approaches, revealing a critical gap in current alignment strategies.

FormalRx transforms opaque autoformalization evaluations into clear, actionable insights, enabling targeted improvements in formal reasoning systems.

A staggering 28% of mid-sized models' answers are mere spurious guesses, exposing critical flaws in their reasoning capabilities.

TimeThink revolutionizes video reasoning by enabling models to pinpoint relevant temporal evidence with unprecedented accuracy, outperforming existing approaches.

Dynamic gating allows DGSeg to filter out noise and ambiguity in segmentation cues, leading to superior performance in reasoning segmentation tasks.

TREK transforms the way models tackle challenging prompts by expanding their exploration support, leading to substantial performance gains even in the hardest task scenarios.

In-process memory access can boost language agent recall performance dramatically while slashing latency by orders of magnitude.

Forged reasoning attacks can completely compromise LLM agents' memory integrity, achieving up to 100% success against existing defenses.

InvWeaver outperforms existing methods by solving 72 out of 82 multi-loop benchmark problems, showcasing a breakthrough in invariant synthesis for complex programs.

dOPSD leverages a model's own decoding process to provide on-policy supervision, leading to substantial gains in reasoning performance without external labels.

LLMs exhibit a surprising mechanism-level routing ceiling, with accuracy jumping by over 10% when provided with specific cues, revealing critical misrouting issues in proof-mechanism classification.

SpecCoder boosts the quality of executable specifications by up to 358% and turns them into active tools for reliable code verification and repair.

Executable vector graphics enable MLLMs to achieve human-like spatial reasoning through a structured visual workspace.

DemoPSD effectively reduces privileged information leakage while enhancing exploration, leading to superior generalization in large language models.

Optimizing stateful evidence navigation in long-context reasoning leads to more effective evidence synthesis and lower distractor retention than traditional methods.

Despite being frontier models, none of the evaluated LLMs could reliably meet critical clinical reasoning standards, with over half of the essential criteria unmet.

Selectively repeating only the most informative tokens can dramatically enhance reasoning in LLMs while slashing computational costs.

RECONTEXT boosts long-context reasoning in LLMs by effectively reusing evidence from the input, leading to superior performance without the need for retraining.

Personalization in language models can significantly alter reasoning paths, leading to substantial drift that may go unnoticed in seemingly fluent responses.

Neural guidance can reduce median conflict counts to zero and dramatically speed up symbolic solvers, but only under specific conditions.

Autonomous research agents can now produce publication-grade manuscripts in frontier physics, achieving substantive findings while navigating complex literature landscapes.

CheckRLM cuts error accumulation in reasoning chains by correcting factual inaccuracies in real-time, outperforming traditional approaches.

APV enables LLMs to discern pedagogical intent with unprecedented accuracy, achieving a correlation of $r=0.958$ with human judgments.

LLM-driven predicate invention can achieve an 80% success rate in ILP, transforming a traditionally expert-dependent process into an automated one.

Visualizing code dependencies can dramatically enhance issue-resolution performance, outperforming traditional text-based navigation methods.

OPSD can backfire, leading to rote memorization instead of enhancing reasoning, but a novel decomposition approach reveals a path to meaningful improvements.

Vision-language models can now achieve remarkable out-of-distribution accuracy by effectively leveraging self-reflection to correct errors in multimodal reasoning.

TUDUM reveals that fine-tuning for Turkish reasoning can enhance linguistic alignment but may compromise benchmark accuracy, challenging assumptions about model performance in multilingual settings.

A$^{2}$utoLPBench offers an endless stream of dynamically generated LP problems, ensuring agents can be tested against fresh challenges without the risk of training data leakage.

Elevating reasoning effort can boost first-try success rates in code generation from 28% to 89%, while adding testing tools fails to enhance reliability.

Iterative refinement of search queries in a continuous latent space leads to a dramatic increase in video retrieval accuracy and reasoning efficiency.

Theoria achieves a remarkable 91.4% precision in certifying AI-generated answers, revealing hidden premises that traditional LLM judges often miss.

Language can be a powerful supervisory tool, enabling imitation learning to outperform traditional methods by providing structured feedback on task performance.

Adversarial training with human demonstrations can significantly enhance the quality and diversity of language model outputs while preserving accuracy.

Direct thread communication in MPLMs cuts context requirements by half, revolutionizing how LLMs tackle complex reasoning tasks.

Refining Cover's theory reveals that low-dimensional data structures can dramatically enhance classification capabilities, challenging traditional assumptions in machine learning.

Calibration-data composition can dramatically enhance quantization performance, with 3.5-bit models outperforming traditional 4-bit baselines by over 20 points.

StochasT reveals that intelligently grouping language tasks can unlock the full potential of LVLMs, enhancing their performance in both single-turn and multi-turn interactions.

CAT enables Large Reasoning Models to intelligently balance efficiency and accuracy, compressing confident responses while thoroughly processing uncertain queries.

Bayesian uncertainty propagation reveals critical failure points in multi-hop reasoning, outperforming traditional methods in complex scenarios.

Existing LLMs fail to maintain internal representations in maze environments, revealing critical limitations in their reasoning capabilities.

VLMs struggle with object-level counterfactual reasoning, achieving only a fraction of human accuracy in spatial tasks.

Overthinking in language models can be curtailed by segment-level credit assignment, leading to a 5.4% accuracy boost on competitive math benchmarks.

PCS narrows the reasoning performance gap between English and other languages in LRM applications, achieving language consistency without heavy resource demands.

Hallucinations in language models often stem from biased inference paths rather than simply missing information, revealing a deeper layer of reasoning failure.

Looping discrete embeddings with continuous hidden states enables near-perfect accuracy in multi-hop reasoning with fewer training steps than traditional methods.

KidnapRAG reveals how a clever sequence of poisoned documents can subvert the reasoning of advanced RAG systems, showcasing a critical vulnerability in their design.

OmniView-Space redefines spatial reasoning in MLLMs, achieving unprecedented accuracy by leveraging dynamic, egocentric evidence mapping.

LRPO achieves superior video anomaly detection by learning from multiple reasoning trajectories without the need for extensive human annotations.

EFlow's innovative separation of evidence retrieval and reasoning processes leads to substantial improvements in long-video reasoning performance, addressing critical biases in existing frameworks.

Memory compression can enhance long-horizon question answering, achieving a 6x increase in evidence-grounded answers while reducing memory usage by over two times.

DART bridges the reasoning gap in zero-shot video temporal grounding, achieving a remarkable 3.5-point mIoU improvement with over 7 times fewer frames.

Retrieval-based visual reasoning can match state-of-the-art performance with half the parameters, revolutionizing how we think about multimodal learning.

GEAR-Seg transforms reasoning segmentation by decoupling perception and deduction, enabling explicit logic tracking and competitive zero-shot performance.

PlanRAG transforms exploratory reasoning by leveraging logical query trees, achieving superior performance in complex query resolution.

Antaeus uncovers hidden logic vulnerabilities by grounding LLM reasoning in the broader context of repository-level code, revealing insights that traditional models miss.

Aha-moment-driven backdoor attacks can redirect reasoning in VLMs while preserving output coherence, making them harder to detect.

Decoupling perception from reasoning in visual tasks leads to a remarkable 93.2% accuracy on V-Star, showcasing a new paradigm for fine-grained visual reasoning.

Graph-native reinforcement learning can boost hypothesis generation in materials science by achieving up to 65% better traceability than traditional models.

AMVL eliminates the critical train-inference mismatch in MLLMs, leading to substantial improvements in reasoning performance across multimodal tasks.

Active-GRPO not only outperforms existing methods in molecular optimization but also redefines how models can adaptively balance imitation and self-discovery during training.

Achieving near-autoregressive accuracy while boosting decoding speed by over 2.4 times could redefine efficiency benchmarks in generative reasoning tasks.

LOTUS achieves a groundbreaking 2.5x-6.9x reduction in reasoning latency while matching explicit chain-of-thought performance at 3B parameters.

Allowing language models to explore unsafe reasoning can actually enhance their ability to discern harmful from harmless prompts, reducing over-refusal without sacrificing safety.

A new approach to reasoning in LLMs reveals that holistic judging can significantly outperform traditional methods, achieving a record 72.9% accuracy on ARC-AGI-2.

Fork-think reduces token usage by 30% and runtime by 57% while achieving superior reasoning performance, challenging the efficiency of traditional methods.

Hard routing in LoRA expert composition can drastically reduce trainable parameters while preserving reasoning performance, challenging the efficacy of soft routing methods.

ISM achieves superior mathematical reasoning performance with 64% and 86% fewer schemas than the best passive baseline, proving that compact, actively maintained memory systems can outperform larger, static ones.

KSP-enhanced CEGAR-tableaux outperforms traditional methods, especially on large satisfiable problems, marking a breakthrough in modal satisfiability techniques.

Evolving principle-guided supervision can boost MLLM reasoning accuracy by up to 24.6%, transforming how we train models for complex decision-making tasks.

Pruning irrelevant visual tokens can boost medical reasoning performance by over 100%, transforming how VLMs approach sparse medical images.

LLMs can exhibit a staggering 68.6% mismatch between their reasoning and final diagnostic outputs, raising serious concerns about their reliability in clinical settings.

Clinically structured rank allocation in BiRG-LoRA boosts medical question answering accuracy while reducing trainable parameters by over 28%.

LLMs show significant vulnerability to logical fallacies, with distinct profiles of resilience that could inform future model training strategies.

CoLT slashes inference time by over 10x while enabling multi-modal models to reason more efficiently with fewer steps.

Achieving an 80.9% success rate in spatial reasoning tasks, RoboSpatialBrain reveals the critical role of selective reasoning activation in embodied AI.

The Relative Surprisal Index reveals that the interplay between token probability and entropy is crucial for optimizing reinforcement learning in language models, leading to substantial performance gains.

By leveraging both UAV and satellite perspectives, SatAgent achieves unprecedented accuracy in spatial reasoning tasks, outperforming existing models by over 25%.

Data referencing errors plague LLMs even in structured tasks, but a lightweight critic model can boost accuracy by up to 12%.

LLMs excel at predicting the direction of change in unfamiliar physics scenarios but falter dramatically in quantitative reasoning, revealing critical gaps in their understanding.

Fixed counterfactual explanations can lead LMs to generate more accurate introspections about their behaviors, even as those behaviors change over time.

A transformer with explicit fuzzy logic not only matches baseline performance but also reveals how it interprets grammatical structures, making model behavior legible.

LASER cuts average latency by up to 38% while only sacrificing 1% accuracy, revolutionizing how we deploy reasoning models on edge devices.

Relying on text-based rationales for dementia classification can actually degrade performance, revealing the need for more sophisticated approaches like DeTAiL.

The same underlying principle governs three seemingly distinct training methods, revealing that the disagreement among answers is the key driver of effective learning.

PixelEyes achieves precise visual localization by separating reasoning from perception, drastically reducing the redundancy in multi-turn visual searches.

Early-stage reasoning failures can be drastically reduced from 64% to 13% with a novel RL approach that penalizes cascading errors in medical VQA.

Higher offline conservatism in training can paradoxically increase vulnerability to reward hacking during online adaptation, challenging long-held assumptions in the field.