#reasoning

40 curated events

papersTODAY 04:00 UTC

NeuroProlog Applies Multi-Task Fine-Tuning to Neurosymbolic Math Reasoning

A revised arXiv paper introduces NeuroProlog, a neurosymbolic approach that pairs language models with symbolic reasoning to improve mathematical problem solving. The authors use multi-task fine-tuning and describe a "cocktail effect," where combining several training tasks yields better results than training on them individually. The work targets a known weakness in LLMs, which often produce fluent but logically inconsistent math solutions.

research/reasoning/neurosymbolic/fine-tuning

papersTODAY 04:00 UTC

Lightning Weave: Capability Composition for More Efficient Reasoning Models

A new arXiv paper introduces Lightning Weave, a method aimed at pushing the accuracy-efficiency frontier of reasoning models. The authors argue that accuracy and inference efficiency often pull toward different reasoning behaviors, making joint improvement difficult. Their approach relies on composing capabilities rather than optimizing the two objectives independently.

reasoning/efficiency/arxiv/llm

papersTODAY 04:00 UTC

Retrieval-Grounded Reasoning Approach Proposed for Universal Multimodal Embeddings

A new arXiv paper introduces a method that grounds chain-of-thought reasoning in retrieved evidence to improve universal multimodal embeddings, which aim to represent text, images and other modalities in one shared space. The authors argue that reasoning steps should be tied to retrieval so that only relevant information shapes the final embedding. The work targets a single model that can handle a range of cross-modal retrieval tasks.

multimodal/retrieval/embeddings/reasoning

papersTODAY 04:00 UTC

arXiv Paper Proposes Formal Theory of Mind Based on Information Access History

A revised arXiv preprint introduces a formal framework for theory of mind that derives an agent's beliefs from the history of information they encountered, rather than assuming those beliefs are already known. The account models factors such as the order of exposure, the source of information, and its perceived credibility. This extends prior formal treatments that generally take beliefs as given inputs.

theory-of-mind/research/reasoning/arxiv

papersTODAY 04:00 UTC

arXiv paper examines deductive, inductive and abductive reasoning in language models

A revised arXiv preprint analyzes how language models handle three forms of reasoning: deduction, induction, and abduction. The authors compare ways tasks are specified to models, such as explicit instructions versus few-shot examples, and argue that current evaluations leave parts of the reasoning picture unresolved. The work is a research paper rather than a product or model release.

research/reasoning/language-models/evaluation

papersTODAY 04:00 UTC

Study Finds Chemical Chain-of-Thought in Reasoning Models Prone to Hallucination

A new arXiv paper examines how language models trained for chemical reasoning use chain-of-thought steps, and finds that the intermediate reasoning frequently contains fabricated content. Testing four reasoning model families across twelve chemistry tasks, the authors report that hallucination is widespread and largely disconnected from the final answer. The work suggests chain-of-thought traces in this domain act more like an unreliable scratchpad than a faithful record of the model's reasoning.

hallucination/reasoning/chain-of-thought/chemistry

papersTODAY 04:00 UTC

T-LoopFormer: Token-Level Elastic-Depth Looped Transformers for Latent Reasoning

A new arXiv preprint introduces T-LoopFormer, a looped Transformer variant that varies how many times each token passes through a shared parameter block rather than applying a fixed depth uniformly. The approach uses dynamic routing to decide token-level elastic depth, aiming to support latent reasoning while keeping parameter counts low. The announcement is an abstract-only submission, so reported results and comparisons are not yet detailed.

research/transformers/reasoning/efficiency

papersTODAY 04:00 UTC

Segment-Aware Listwise Alignment Targets Reasoning Safety in Large Reasoning Models

A new arXiv paper argues that safety alignment for large reasoning models must address two surfaces at once: the intermediate chain of thought and the final answer. The authors note that existing methods typically align whole responses, which can leave harmful reasoning steps intact even when the visible answer looks safe. Their proposed approach, segment-aware listwise alignment, treats reasoning traces and outputs as distinct segments to be optimized together.

safety/alignment/reasoning/research

papersTODAY 04:00 UTC

TimeThink: Method Aims to Improve Compositional Reasoning in Time-Series LLMs

A new arXiv paper introduces TimeThink, a technique intended to help time-series multimodal large language models reason more compositionally. The authors note that such models often struggle to capture dynamic temporal patterns when answering questions. The work focuses on eliciting stronger reasoning behavior from these models rather than treating forecasting as pure pattern matching.

research/time-series/reasoning/multimodal

papersTODAY 04:00 UTC

Func-R1: Method Aims to Improve Mathematical Function Reasoning in Multimodal LLMs

A new arXiv paper introduces Func-R1, an approach aimed at strengthening mathematical function reasoning in multimodal large language models. The work targets the challenge of combining visual perception with symbolic logic when solving math problems from images. The abstract frames deliberate mathematical reasoning in visual settings as an indicator of advanced multimodal model capability.

multimodal/reasoning/math/llm

papersTODAY 04:00 UTC

Study Examines Training Domain Specialists Without Reasoning Trajectories

A new arXiv paper looks at domain expert distillation, where a teacher model normally passes reasoning traces to a student model. It investigates what happens when specialists are trained only on question-answer pairs and no explicit reasoning supervision is provided. The work aims to clarify how much the reasoning trajectories actually contribute to the student's performance.

distillation/reasoning/fine-tuning/arxiv

papersTODAY 04:00 UTC

VeriDx framework verifies clinical diagnoses through disease-centric obligations

Researchers propose VeriDx, a verification approach for clinical reasoning that ties each disease hypothesis to obligations such as checking key evidence, ruling out alternatives, and resolving contradictions. The method aims to distinguish diagnoses reached through sound reasoning from those that are correct by coincidence. It is described in an arXiv preprint (2609.14018v1).

clinical-ai/reasoning/verification/arxiv

papersTODAY 04:00 UTC

arXiv paper studies how evidence ordering shapes streaming test-time compute

A new arXiv paper argues that reasoning policies for a fixed task and compute budget should change depending on the order in which evidence arrives. It frames this as an "information-slack dilemma": computing early leaves more time to finish but relies on incomplete evidence, while waiting yields better information at the cost of fewer remaining steps. The authors contend that depth and width alone do not capture this tradeoff, which streaming settings must account for.

test-time-compute/reasoning/arxiv/llm

papersTODAY 04:00 UTC

Paper studies switching between language and symbolic forms for spatial reasoning

A revised arXiv preprint examines how reasoning improves when models move between natural language and symbolic representations such as grids or sketches. The authors argue that human problem-solving is multimodal, with people offloading difficult steps into diagrams to expose structure and reduce errors. The work proposes treating this modality shift as a mechanism AI systems can adopt for spatial tasks.

reasoning/multimodal/arxiv/spatial-reasoning

papersTODAY 04:00 UTC

DARE: Dialectical Agentic Reasoning for Structured Knowledge Fact Checking

A new arXiv paper introduces DARE, a method for verifying natural language claims by reasoning over structured evidence such as knowledge graphs. It builds on program-generation approaches in which large language models produce executable graph reasoning routines, adding a dialectical agentic layer to improve verification. The work targets fact checking tasks where evidence is organized in structured form rather than plain text.

fact-checking/agents/knowledge-graphs/reasoning

papersTODAY 04:00 UTC

Teacher-Guided Curriculum Boosts Data Efficiency in RLVR Training

A new arXiv paper addresses a known failure mode in reinforcement learning with verifiable rewards (RLVR), where training problems that are too hard for a model produce uniformly failed attempts and yield no learning signal. The authors propose a teacher-guided curriculum that sequences training data so the model encounters problems it can actually solve, making the process more data-efficient. The work targets mathematical reasoning in large language models and falls within the cs.CL area.

rlvr/reinforcement-learning/curriculum-learning/reasoning

papersTODAY 04:00 UTC

arXiv paper critiques on-policy self-distillation, proposes RL contrastive method

A revised arXiv paper examines on-policy self-distillation, a technique that gives reasoning models dense token-level feedback by matching their output distribution to one produced with extra context such as a verified solution. The authors argue this approach yields a flawed distribution, and they introduce RLCSD, which combines reinforcement learning with contrastive self-distillation on policy. The work is a preprint and has not been peer reviewed.

self-distillation/reinforcement-learning/reasoning/arxiv

papersTODAY 04:00 UTC

Paper asks whether reasoning models show systematicity on rule induction tasks

A new arXiv preprint examines whether AI reasoning models generalize systematically, meaning that learning one concept should transfer to closely related variations of it. The authors evaluate these models on rule induction tasks to test how robustly that kind of structured generalization holds. The work questions whether current reasoning capabilities reflect genuine systematic understanding or narrower pattern matching.

research/reasoning/evaluation/generalization

papersTODAY 04:00 UTC

arXiv paper examines how LLMs perform propositional logical reasoning

A revised arXiv preprint (2601.04260v2) investigates the internal computations LLMs use when solving propositional logic tasks. The authors argue that earlier mechanistic interpretability work focused on task-specific circuits, leaving broader questions about the underlying computational structure unanswered. The paper appears in both cs.AI and cs.LG listings as a replacement submission.

interpretability/reasoning/llm/arxiv

papersTODAY 04:00 UTC

CapGeo-Bench separates visual perception from geometric reasoning in multimodal models

A new arXiv paper introduces CapGeo-Bench, a benchmark designed to evaluate geometric understanding in multimodal large language models while distinguishing failures in visual perception from failures in reasoning. The authors note that even strong closed models such as GPT-o3 continue to lag on geometry problems despite success on purely textual math tasks. The benchmark aims to give a clearer picture of where these systems break down.

benchmarks/multimodal/reasoning/geometry

papersTODAY 04:00 UTC

Paper Distinguishes Difficulty Detection From Deliberation Allocation in Reasoning Models

A new arXiv paper examines how large reasoning models spend extra computation on problems that humans also find harder, noting that longer traces do not necessarily mean the model is deliberately allocating more effort. The authors split the concept into two parts: registering that a problem is difficult, and deciding how much additional reasoning to devote to it. This framing is intended to clarify when extended reasoning reflects genuine deliberation versus a byproduct of problem difficulty.

reasoning/research/llm/arxiv

papersTODAY 04:00 UTC

Verifier-Gated Multi-Expert Distillation Aimed at Scientific Reasoning

A new arXiv paper examines multi-teacher on-policy distillation, the technique of training specialist models and then transferring their abilities to a single student using the student's own generated outputs. The authors propose assigning supervision token by token rather than sequence by sequence, with a verifier deciding which expert teacher should guide each token. The method is aimed at scientific reasoning tasks.

distillation/reasoning/multi-expert/arxiv

papersTODAY 04:00 UTC

arXiv Paper Examines Argument Structure and Proof Methods Across Genres Using LLMs

A new arXiv preprint in computational linguistics studies how arguments are structured when a direct proof of a claim is difficult, and how an alternative but related statement can be used instead. The authors apply large language models to compare argumentation and proof patterns across different text genres. The work falls within NLP research on reasoning and argument mining.

research/reasoning/llm/argumentation

papersTODAY 04:00 UTC

Paper studies how distractors affect test-time scaling in reasoning VLMs

A new arXiv preprint examines whether irrelevant information, known as distractors, changes how vision-language models behave when allowed to spend more compute at inference time. Prior work on text-only models found that such distractors can worsen inverse scaling, where reasoning degrades as test-time compute grows. The authors extend that question to multimodal settings, where models must handle both images and text. The submission is a cross-listed replacement in cs.AI and cs.LG.

vision-language/reasoning/test-time-scaling/arxiv

papersTODAY 04:00 UTC

Universe of Universes Framework Treats LLM Ecosystem as Retrieval Corpus

A new arXiv preprint proposes the Universe of Universes framework, which models the collection of major large language models as a structured corpus that can be queried rather than as isolated systems. The authors describe a compositional architecture combining automated reasoning and machine learning, and introduce concepts such as benefit yield functions, implosion thresholds, and infrastructure-aware optimization for coordinating multiple LLMs.

multi-llm/reasoning/optimization/arxiv

papersTODAY 04:00 UTC

arXiv paper proposes skill-augmented graph reasoning for table question answering

A revised arXiv preprint introduces a method for table question answering that treats questions differently instead of uniformly, pairing learned skills with graph-based reasoning over table structures. The authors argue that reporting only overall accuracy hides a sharp divide between easy lookup questions and harder multi-step operations. The approach, called skill-augmented table graph reasoning, targets operation-wise evaluation of large language models on tabular data.

table-qa/reasoning/llm-evaluation/arxiv

papersTODAY 04:00 UTC

Deliberative Diagnostic Framework for Evaluating LLM Opinion Simulation

A new arXiv paper proposes a diagnostic method for testing whether large language models genuinely reason about newly presented information or simply reproduce opinions they absorbed during training. The authors argue this distinction matters for "silicon sampling," where LLM personas stand in for human respondents in public-opinion research. Existing evaluations, they note, only check whether a simulated persona gives a plausible answer rather than how it handles fresh evidence.

llm-evaluation/reasoning/silicon-sampling/research

papersTODAY 04:00 UTC

arXiv Paper Proposes White-Box Meta-Prompting Method for Adaptive RAG and Reasoning

A new arXiv preprint introduces CWM, a controllable white-box meta-prompting approach aimed at making retrieval-augmented generation and reasoning more adaptive in large language models. The work targets the problem of deciding how and when models should draw on external knowledge alongside their own reasoning. The abstract is brief, so full evaluation details remain unclear.

rag/meta-prompting/reasoning/arxiv

papersTODAY 04:00 UTC

arXiv Paper Proposes 'Post-Reasoning' Method to Improve Non-Thinking LLMs

A revised arXiv preprint introduces a technique called post-reasoning that aims to raise the quality of large language model outputs produced without intermediate reasoning traces. The work targets the cost and latency that come from generating long chains of thought at inference time, arguing that many practical tasks do not need them. The authors present the approach as an add-on that improves non-reasoning models without extra expense.

reasoning/inference/efficiency/llm

papersTODAY 04:00 UTC

arXiv Paper Diagnoses and Improves Visual Chain-of-Thought for Geometry Solvers

A revised arXiv preprint argues that multimodal models need active visual assistance, such as drawing auxiliary lines, to handle complex geometry problems. The authors examine shortcomings in current evaluation of visual chain-of-thought methods and propose ways to strengthen how models reason with diagrams. The work falls under cs.AI and focuses on diagnosing and improving these visual reasoning pipelines.

multimodal/reasoning/geometry/benchmarks

papersTODAY 04:00 UTC

arXiv Paper Examines Capacity Limits of Reasoning via Superposition

A new arXiv preprint studies how much intermediate computation a single vector can carry when language models reason through continuous or recurrent methods rather than token-by-token chain-of-thought. The work frames multi-step reasoning as superposition, where partial computations are packed into hidden states, and analyzes the resulting capacity limits. It offers a theoretical lens on the trade-offs between explicit token-based reasoning and continuous latent approaches.

research/reasoning/superposition/llm

papersTODAY 04:00 UTC

Stellar Colosseum: a multi-agent harness for long-horizon math and TCS research

Researchers posted an arXiv preprint describing Stellar Colosseum, a model-agnostic framework that coordinates multiple language-model agents on extended research problems in mathematics and theoretical computer science. The authors argue that while models can generate convincing short proofs, they remain unreliable when progress requires many uncertain, interdependent decisions in sequence. The harness is presented as a way to structure such long-horizon work rather than a single model release.

agents/research/reasoning/arxiv

papersTODAY 04:00 UTC

arXiv paper formalizes and measures abductive 'jump' in large language models

A new arXiv preprint proposes a formal definition and measurement method for the 'jump' — the abductive step from evidence to a new set of axioms — in large language models. The work responds to ongoing debate over whether LLMs are structurally incapable of such reasoning, framing the question as something that can be tested empirically rather than argued abstractly.

research/reasoning/llm/evaluation

papersTODAY 04:00 UTC

On-Policy Self-Distillation Method Aims to Prevent Entropy Collapse in RL-Trained LLMs

A new arXiv paper proposes an approach called on-policy self-distillation that acts as a "policy reheater" for reinforcement learning with verifiable rewards. The authors target entropy collapse, a failure mode where model policies become overly concentrated, cutting rollout diversity and weakening the learning signal. The method is presented as a way to keep exploration alive during RL training of large language models.

reinforcement-learning/llm-training/reasoning/arxiv

papersTODAY 04:00 UTC

FinAcumen agent uses self-evolving experience memory for financial multimodal reasoning

A new arXiv paper introduces FinAcumen, an agent design for financial tasks that must combine arithmetic, information retrieval, chart and document interpretation, and time-based reasoning over mixed evidence. While earlier tool-using agents focused mainly on executing steps accurately, this work targets the remaining difficulty of handling heterogeneous inputs, proposing a memory harness that evolves from its own accumulated experience. The submission is a revised version of the preprint.

agents/finance/multimodal/reasoning

papersTODAY 04:00 UTC

KREL Method Uses LLM Reasoning Over Clinical Evidence for Automatic Medical Coding

A revised arXiv paper presents KREL, a technique for automatically assigning standardized ICD codes to clinical notes. The method relies on knowledge-guided reasoning over clinical evidence using large language models, aiming to support reimbursement, quality reporting and research. The submission is a replacement version of an earlier preprint.

llm/healthcare/reasoning/medical-coding

modelsTODAY 12:00 UTC

Salesforce and Nvidia release Koa reasoning model for business tasks

Salesforce introduced Koa, a reasoning model developed with Nvidia and built on Nvidia's open-weight Nemotron foundation. The model is tuned for enterprise functions such as sales, marketing and customer support rather than general-purpose use. The companies position it as an alternative to models from the major AI labs.

salesforce/nvidia/open-weight/reasoning

papersSEP 11 04:00 UTC

Paper Measures AI Progress Toward Mathematical Discovery with Automatic Verification

A revised arXiv preprint introduces a method that uses automatic verification to track how well language models reason about unsolved mathematical problems. The author notes that although large language models now handle sophisticated math and science reasoning, whether they can contribute genuinely new research remains contested and thinly studied. The work aims to give a measurable way to assess progress on that question.

research/reasoning/benchmarks/math

papersSEP 11 04:00 UTC

RecurTrace pairs latent recurrent reasoning with adaptive loop-time memory

A new arXiv paper examines how repeating a small block of middle layers can raise a language model's effective inference depth without extra parameters or additional generated tokens. The authors propose RecurTrace, which uses adaptive latent reasoning and a loop-time memory mechanism to address two design choices in such recurrent architectures. The work is a revised preprint and reports on improving reasoning through latent recurrence.

research/reasoning/llm-architecture/latent-recurrence

papersSEP 12 13:27 UTC

Study links reasoning models' internal states to distinct thought steps

A new study finds that operations such as arithmetic, recalling formulas, and logical deduction show up as separate patterns inside reasoning models, most visibly in their middle layers. This suggests models carry out more processing than their published chain-of-thought text discloses, which researchers flag as relevant to AI safety and oversight. The findings could inform how developers monitor or audit model reasoning.

interpretability/reasoning/safety/research