Reasoning & Chain-of-Thought - Weekly Roundup

Realtime-Venus, a proactive full-duplex interaction system with two separately trained 9B models: Realtime-Venus-Omni for audio-visual interaction and Realtime-Venus-Audio for spoken interaction, which leads the compared models on MMAU, Llama Questions, and Speech CMMLU.

StepAudio 3 Realtime, an audio-language foundation model organized around a continuous listen-converse-think-act loop, achieves dialogue and reasoning performance comparable to dedicated reasoning models while speaking in real time, and resolves the tension between deep deliberation and latency via Think-While-Speaking.

This work shows that specialist optimization implicitly selects from this latent trajectory space, and establishes a new view of specialist training: when gold reasoning is absent, tuning choices directly control the latent supervision passed to downstream models.

The results suggest that fine-tuning is not just about how much a model changes, but how that change is spent, and that changing the accessible directions can qualitatively alter the outcome of fine-tuning.

Realtime-Venus, a proactive full-duplex interaction system with two separately trained 9B models: Realtime-Venus-Omni for audio-visual interaction and Realtime-Venus-Audio for spoken interaction, which leads the compared models on MMAU, Llama Questions, and Speech CMMLU.

StepAudio 3 Realtime, an audio-language foundation model organized around a continuous listen-converse-think-act loop, achieves dialogue and reasoning performance comparable to dedicated reasoning models while speaking in real time, and resolves the tension between deep deliberation and latency via Think-While-Speaking.

This work shows that specialist optimization implicitly selects from this latent trajectory space, and establishes a new view of specialist training: when gold reasoning is absent, tuning choices directly control the latent supervision passed to downstream models.

The results suggest that fine-tuning is not just about how much a model changes, but how that change is spent, and that changing the accessible directions can qualitatively alter the outcome of fine-tuning.

This work introduces StepAudio 3 Music, a large-scale, long-form music generation model that supports explicit musical planning and open-domain text-controlled generation, and achieves the highest AudioBox Content Enjoyment, Content Usefulness, and Production Quality scores and the highest MuQ-MuLan similarity among the evaluated systems.

Across reasoning, long-context understanding, and agentic tasks, SAS consistently outperforms trainable sparse attention baselines across attention budgets, with especially large gains under tight budgets, demonstrating more effective context ranking for downstream tasks.

This work presents ZGCM-1, a fully open 7B dense foundation model trained from scratch with extreme data, system, and algorithmic efficiency, and establishes an AI-native R&D workflow where agent swarms autonomously manage cluster operations, data curation, and rapid diagnostic evaluation.

It is demonstrated that a general-purpose agent can directly drive a physical robot throughout task execution without any task-specific or environment-specific training and Agent as Policy (AGP) is introduced, which places task planning and execution under the agent's control.

Generative Verification (GenV) is introduced, which distills an offline Z3-equivalence oracle into a reference-free, continuous reference-equivalence score by repurposing the language model's native vocabulary space and theoretically proves that structural, verdict-only verification heuristics are mathematically bounded to chance-level detection on these deceptively valid traces.

The Feature Recovery Module (FRM), a plug-and-play module that maps degraded encoder features to pristine-aligned representations while keeping the host frozen, is proposed, which improves scene-level detection, CLIP/SigLIP2 feature recovery, and all four object-level VLM tasks, with larger gains under more severe degradation.

This work shows that for every fixed rational discount factor ($\gamma\in(0,1)$), exact planning remains NP-hard, and introduces a randomized polynomial-time approximation scheme for every fixed look-ahead depth.

SIRF (Spec-Internalized Risk Foundation Model), which internalizes a platform's complex policies, synthesized without additional human annotation via EntiGraph, MAGA rewriting and account-level chain-of-thought (CoT), into the weights via continued pretraining (CPT), so rules are applied at high precision under an ultra-low-latency, verdict-only deployment.

Evaluation on a representative multimodal benchmark demonstrates that EMMI can reduce the communication payload by 32x while maintaining comparable downstream accuracy, resulting in up to a 3.4x reduction in estimated end-to-end inference latency under bandwidth-constrained edge conditions.

Ecdysis is proposed, an efficient and effective framework that distinguishes model-specific accommodation from harness-level repair and biases adaptation toward systematic harness deficiencies by identifying recurring cross-task failure patterns.

Results show that adaptive direction can emerge without being explicitly specified as a behavioral objective, and the same persistence that makes such adaptive agency useful can also allow misalignment, corrupted state and unintended behavior to persist across task boundaries.

This paper introduces Magenta, a training-free agentic pipeline that, given only a natural-language problem, produces an answer, expresses it as a Lean 4 statement, and constructs a machine-checked proof, which achieves 100% accuracy across all evaluated olympiad benchmarks.

HALDETECT, the authors' system for the English hallucination-detection track (Task 1b) of ImageEval 2026, in which a system must identify, from an image and three culturally plausible statements, the single visually grounded one, is presented.

This work proposes Vid-PRE (Video Prompt Reasoner and Enhancer), a model-agnostic prompt rewriter that offloads the cognitive burden of reasoning to a dedicated VLM, and introduces VWG-Bench, a comprehensive benchmark spanning 9 reasoning dimensions and 38 fine-grained tasks.

ChurnBench is presented, an open-source benchmark that generates a four-source enterprise data fabric as a timeline rather than a snapshot, and it is found that when a system refreshes on a schedule, cache age does not predict staleness.

Empirical results show that CamPilot outperforms state-of-the-art text-to-movie generation methods on cinematographic control and quality, highlighting the impact of professional camera design on movie generation.

EgoGenEval introduces EgoGenEval, a geometry-grounded, pose-free benchmark designed to evaluate the physical consistency of visual generators under ego-motion, and shows that pairwise supervision does not reliably improve camera-motion grounding and scene-state preservation together.

OststaDiff is presented, a decomposed diffusion-policy framework with a lightweight obstacle-aware visual encoder that extracts a structured target-obstacle-background representation, enabling the downstream alignment policy to generate end-effector trajectories toward a target-centered bottleneck pose while reasoning about surrounding obstacles.

These results validate the deployed navigation and inspection closed loop, while HROS provides an extensible software foundation for memory-augmented, voice-aware, and continuously improvable embodied inspection agents.

LASP is introduced, a Language-Augmented Semantic Priors framework that leverages large language models to infer structured, solver-usable B-spline priors from procedural modeling histories, establishing a new paradigm of language-guided geometric optimization in modern CAD systems.

BlueSTAR is presented, a tiered agentic architecture for autonomous cyber defense in enterprise IT/OT networks that retains the fast containment of deterministic response for known threats while successfully defending against attacks requiring contextual and cross-cycle reasoning.

Negative Self-Distillation (NSD) is introduced, a new framework that optimizes LLMs by diverging from flawed reasoning rather than imitating privileged solutions, and consistently outperforms OPSD and other label-free, self-bootstrapping reinforcement learning (RL) baselines.

CausalArena is introduced, a unified and evolvable benchmark for causal discovery under a common protocol, and substantial ranking shifts across SCM families and protocols are revealed, showing that strong performance in one benchmark regime does not reliably transfer to others.

Looped flows are proposed, an approach that allows solving harder problems by spending more computation through a finer temporal grid and enables multiple valid predictions from different initial noise samples.

Despite its weaker standalone accuracy, the reverse posterior serves as a more useful anchor than forward-only alternatives, providing complementary information for collective decision-making.

This work proposes Test-Time Adaptation via Stability-Aware Confidence Optimization (TASCO), a framework that incorporates local stability into confidence-based test-time adaptation while keeping the LLM frozen and improves reasoning accuracy and token efficiency across diverse LLMs and reasoning benchmarks.

Sci-MMR is introduced, a benchmark for multi-step evidence-grounded scientific reasoning built on structured argument graphs linking scientific claims, citation-grounded knowledge, visual evidence, and supporting regions, and it is found that current answer-centric benchmarks substantially overestimate the evidence-grounded reasoning capabilities of multimodal research agents.

Whether vision language models can decide when to answer immediately and, when more evidence is needed, which experiment to perform is studied to reveal limitations in physical reasoning that conventional answer accuracy can overlook.

This work introduces MindTopo, a benchmark of topological intuition across five properties grounded in cognitive science and formal topology: continuity, separation, order, enclosure, and knots, which performs better on reasoning than on planning, and the best-performing model remains far below observed human performance.

A Database Normalization Benchmark (DNBENCH), comprising 3,275 samples for evaluating LLM-driven database normalization from 1NF to BCNF, and proposes Multi-Agent Reasoning for Schemas (MARS), which separates evidence extraction, violation diagnosis, and decomposition planning from schema generation and verification.

The proposed calculus gives a common setting for CDCL(T)-style SMT solving, a range of instantiation-based procedures, and non-ground clause learning, and it is proved that it simulates CDCL, SCL(FOL), SCL(T), and even Resolution.

Across answer-focused, mixed-reasoning, and CoT-sensitive benchmarks, routed control improves robustness over fixed long decoding, pure short decoding, and single-rule interventions.

This paper presents SurgicalRoomAgent, a voice-interactive multi-agent system for smart operating rooms based on large language models (LLMs), which achieves natural language understanding, device control, intraoperative recording, and surgical report generation through a layered architecture comprising a voice interaction pipeline and an agent core.

Xiaomi-CocktailASR-1 achieves state-of-the-art performance, effectively addressing the cocktail party problem through a unified architecture that balances multispeaker and single-speaker recognition accuracy, along with rejection capability.

Structurally Speaking is introduced, a lightweight structured prompting protocol that guides translation between explicit connectivity and motif-level abstraction and suggests that explicit topology-to-motif reasoning guidance can make LLM-generated graph captions more interpretable without model fine-tuning.

Evaluated on recommendation models, Auto-RecSys significantly reduces the human time required per experiment cycle and improves execution reliability as its playbooks mature.

EHR (Emotion Hallucination Rate), an evaluator that quantifies emotion hallucinations across six facets, and HMER (Hallucination-aware Memory-guided Emotion Reasoning), a training-free framework for emotion hallucination mitigation, which enables fine-grained mitigation across diverse hallucination facets.

Benchmark Radar is presented, a living database and search engine for retrieval and discovery of AI benchmarks, covering LLM evaluation, agentic and tool-use benchmarks, coding, reasoning, safety, and domain-specific evaluations.

It is shown the path to generalizing L2 reasoning to held-out languages goes through broader language coverage, readily available multilingual non-reasoning data, and a sufficient English reasoning backbone, indicating that reasoning is a language-agnostic behavior that can be transferred across typologically diverse languages through careful data mixing.

All three AI chatbots exhibited a high level of ability in the understanding of radiological cases with abnormal anatomy, indicating their potential use as a supplementary tool in radiological assessment and radiological diagnostic support.

Model scale is no longer the sole bottleneck for complex causal analysis: pairing RL with simulator-synthesized interventions allows a 35B parameter model to surpass Claude Opus 5 on real-world diagnostic tasks.

Multi-source Bayesian network fusion usually destroys inference tractability, but imposing hard treewidth constraints via evolutionary search preserves consensus graph structure without sacrificing computational feasibility.

Automated micro-step prompt refinement can more than double an agent's domain-specific reasoning and scientific rigor scores without fine-tuning underlying model weights.

Optimizing against frozen reward models collapses true executed reasoning performance by 90% under GRPO, but feeding back an on-policy stream of just 10% reality-settled labels closes the hacking gap and preserves 6x the reward.

SHAP can explain why a customer might leave, but combining cost-aware counterfactual scoring with LLM action retrieval yields an 80% greater predicted churn-risk reduction without introducing factual profile hallucinations.

Process supervision does not require generating explicit text: continuous latent reasoning can halve token generation costs while outperforming standard CoT through soft prototype alignment.

Instead of mimicking gold-standard examples, MLLMs reason significantly better when shown suboptimal outputs alongside explicit contrastive paths to correct them.

Full rationale supervision burns up to 254× more tokens without guaranteeing reliability, whereas isolating representations with high rationale-boundary sensitivity selectively immunizes medical LLMs against choice-order brittleness.

Forcing a behavioral model to generate a written rationale before predicting an outcome degrades ranking accuracy by up to 14.5 AUC points compared to directly reading out answer token probabilities.

Standard SFT wastes gradient budget on tokens models either already know or cannot yet grasp; trimming supervision from both extremes yields up to a +26.9 point boost on MATH500 with zero reference-model overhead.

Gating multi-teacher distillation on candidate correctness induces catastrophic label collapse and zero minority-class recall while failing to outperform simple hard filtering or improve factual grounding.

Standard SFT degrades financial reasoning benchmarks by up to 4 percentage points, but pairing self-distillation with rule-verified GRPO transforms domain post-training from a capability tax into a 3-point gain.

Stopping reasoning models early via intermediate answer consensus routinely truncates self-correction, because repeated agreement on a partial trajectory reflects prompt persistence rather than completed reasoning.

This work introduces MetroLLM-Bench, a 955-case benchmark for testing language models as the policy layer of a transit kiosk, and evaluates twenty-six models from six vendors, of which twenty-three are ranked.

A simple, training-free framework in which a reasoning-capable vision-language model iteratively searches and reasons over Wikipedia, gathering evidence dynamically, shows that reasoning and retrieval are complementary on rare entities.

An open-model test-time-compute pipeline for natural-language proof generation for hard olympiad mathematics that operates entirely in natural language, with no formal prover, external tools, or internet access is presented.

Nearly 40% of GRPO rollouts yield zero gradient due to uninformative all-or-nothing groups, but an offline anchor pass can slash this cold-start compute waste before the target policy generates a single token.

Strict physical conservation guarantees over LLM reasoning graphs achieve machine-verifiable path auditing down to machine epsilon, even as pre-registered falsification shows zero accuracy advantage over trivial keyword heuristics on real-world benchmarks.

Reasoning traces with identical final answers and identical entropy curves can conceal radically different internal dynamics, from protracted multi-hypothesis competition to sudden late-stage revision.

The structured reasoning traces users prefer actively degrade oversight, triggering higher false-alarm rates and unearned trust compared to plain chain-of-thought.

Continuous diffusion fails at discrete reasoning primarily because standard reverse samplers stubbornly anchor to early mistakes: simply jumping directly to clean predictions spikes Sudoku validity from 31% to 95% with zero retraining.

Length generalization in transformers does not require double-exponential sample sizes: viewing inputs as compressed strings via power words slashes worst-case sample complexity down to polynomial.

The choice of elementary Boolean connectives draws a sharp, fully classified line between polynomial-time PAC learnability and intractable empirical risk minimization.

Domain gaps baked in during mid-training are virtually immutable: compensatory SFT failed to bridge a single domain performance disparity at a 5% threshold despite lifting overall baseline accuracy.

This is the first end-to-end demonstration that a seven-person independent team can train open-weight models with leading agentic cyber capability and all three checkpoints rank 1st among models at comparable parameter scales.

Deeper reasoning creates an unexpected safety backdoor: extended CoT dilutes attention away from system constraints, making models progressively more vulnerable to jailbreaks the longer they "think."

Standard semantic triples collapse all negative knowledge into a single binary absence, blinding structured representations to the critical distinctions between contradictory, opposite, and intermediary states required for counterfactual reasoning.

Static textbook circuit problems can now be autonomously transformed into causally grounded, interactive learning environments by coupling multimodal diagram parsing directly with deterministic physics engines.

This work introduces SCHEMEARENA, a 400-scenario benchmark for scalable scheming stress testing, constructed through a factorized scenario synthesis framework spanning diverse safety-relevant tool domains, instrumental goals, oversight conditions, and pressure mechanisms.

The theoretical speedup of $k$-tuple lattice sieving faces fundamental geometric limits, with its maximal rate now pinned to an asymptotic band of $\Theta(\frac{\log k}{k})$ within a factor of two.