arXiv:2605.23071v2 Announce Type: replace Abstract: Large language models (LLMs) increasingly rely on long-context processing, but expanding context windows introduces substantial computational and financial costs. Existing context reduction approaches, including retrieval and memory compression methods, are typically evaluated using performance and efficiency metrics independently, limiting systematic comparison and deployment-aware decision-making. This paper introduces The Efficiency Frontier, a unified framework for cost--performance optimization in LLM context management. The framework models context strategy selection as a deployment-aware optimization problem that jointly accounts for task performance, token cost, and preprocessing reuse through amortized cost modeling. Unlike existing evaluations that compare methods in isolation, the proposed framework enables decision-oriented analysis. It identifies when different context management strategies become preferable under varying operational conditions. Experiments on HotpotQA reveal distinct operational regimes and transition boundaries between retrieval-based and preprocessing-based strategies. Results show that deployment-aware optimization reduces effective token usage by approximately 25% at comparable performance, enabling more cost-efficient deployment of large language model systems, while amortized memory compression achieves over 50% lower token cost relative to full-context prompting in higher-performance settings. Overall, the proposed framework provides a principled and practical foundation for evaluating and deploying scalable, efficient, and sustainable LLM systems across enterprise, scientific, and public-sector applications.
Science Journals
arXiv:2605.23243v5 Announce Type: replace Abstract: We evaluate whether frontier LLMs are ready for cybersecurity through a dual-mode benchmark: white-box function-level vulnerability detection (VulnLLM-R, across C/Java/Python) and black-box web application security testing (five production-style applications with 118 ground-truth vulnerabilities across 20+ CWE families, which we will open-source). We test six frontier models (GPT-5.4, Codex~5.3, Claude Opus~4.7, Sonnet~4.6, Gemini~3.1~Pro and Gemini~3~Flash) and two domain-specialized models across four testing paradigms. Our findings are sobering: (1)~every frontier model produces 10-50% false positive rates in white-box detection, systematically over-predicting vulnerabilities; (2)~in black-box testing, frontier models achieve only 4-8% ground-truth coverage, improving to just 10-19% even with external security tools (Playwright MCP, Burp Suite MCP); (3)~structured penetration-testing methodology encoded in domain-specialized agents raises per-family detection above 50%, demonstrating that methodology, not scale, is the primary lever; and (4)~a domain-specialized defense model achieves the highest precision (0.904) and lowest false positive rate (9.7%) among all models, on a single GPU. We identify the absence of structured security testing traces end-to-end request/response sequences, failure-heavy data, and multi-step attack chains as the fundamental training data bottleneck, and propose self-play security testing as a data generation strategy. Our results make the case for vertical foundation models purpose-built for cybersecurity.
arXiv:2605.23660v2 Announce Type: replace Abstract: The rapid advancement of Large Language Models (LLMs) has introduced new possibilities and challenges in physics education, necessitating rigorous evaluation of their capabilities as both problem solvers and automated assessors. This paper presents the results of three complementary studies that evaluated frontier models released between mid-2024 and late-2025. Models were assessed on their ability to generate accurate, step-by-step solutions to university-level physics problems in Classical Mechanics, Electromagnetism, and Quantum Mechanics, and subsequently on their reliability in grading student solutions against a formal mark scheme. The results indicate a clear trajectory toward benchmark saturation in text-based reasoning, with recent architectures (such as ChatGPT-5.1 and Gemini 3.0 Pro) achieving near-perfect scores. Furthermore, recent advances in native multimodal integration have resolved previous limitations in spatial geometry and topological interpretation, enabling models to accurately process accompanying diagrams. As automated assessors, newer models demonstrated significant improvements in alignment with human grading, heavily mitigating the systemic over-marking observed in earlier iterations. However, while models reliably evaluate fully correct handwritten work, assigning partial credit to flawed or incomplete reasoning remains a persistent challenge. These findings suggest that as of late 2025, LLMs offer viable support for both independent student learning and instructional automation, provided their limitations in evaluating ambiguous reasoning are actively managed.
arXiv:2605.24844v2 Announce Type: replace Abstract: While general-purpose Large Language Models (LLMs) applied to Geology often hallucinate when reasoning about subsurface structures and deep-time evolution, current AI in Earth sciences predominantly targets surface remote sensing and GIS. To bridge this gap, we introduce Geo-Expert, a family of parameter-efficient geological LLMs fine-tuned on a custom-curated, high-quality instruction dataset processed using our custom instruction synthesis pipeline. We investigate the impact of model scaling and architecture by fine-tuning three base models: Qwen3-8B, Qwen3-32B, and Gemma-3-27B, with Low-Rank Adaptation (LoRA) method. Our extensive evaluation on a novel domain-specific benchmark, Geo-Eval, reveals that a domain-aligned 8B model can outperform open-weight 70B generalists and proprietary GPT-4o on specialized geological reasoning, while a 32B variant approaches frontier reasoning models. The optimized 8B model further offers a competitive cost-performance ratio for deployment. This work provides a reproducible recipe for democratizing scientific LLMs and establishes a baseline for geological artificial intelligence.
arXiv:2605.24896v3 Announce Type: replace Abstract: Seasonal forecasting of summer rainfall in East Asia remains a grand challenge, as predictability at 3 to 6 month lead times is constrained by the spring predictability barrier, weak large-scale signals, and localized nonlinear convective extremes. We address this challenge with CAPES, which integrates a kilometer-resolution coupled regional model with atmosphere, land, and ocean components and a data-driven AI seasonal forecasting system. At 15 km resolution, the fused workflow combines 174 numerical members from varying start times, physics schemes, and parameter perturbations with 1,600 AI members generated from initial and physical perturbations. Using the full LineShine system, CAPES completes ten annual 1,774-member hindcasts for 2016 to 2025 within 14.6 hours, improving the mean prediction score from ECMWF's 71.8 to 75.9 and delivering a major gain in operational forecasting capability. The 1-km configuration further enables fine-scale typhoon simulation and establishes the feasibility of kilometer-scale fused ensemble forecasting on a one-week timescale.
arXiv:2605.25044v2 Announce Type: replace Abstract: Learning universal policies from cross-embodied data remains a fundamental challenge in robotics. Although Vision-Language-Action (VLA) models are pre-trained on large and diverse datasets, they typically rely on embodiment-specific fine-tuning to achieve strong performance in downstream tasks. This requirement severely limits their generalization capability and restricts knowledge transfer across embodiments performing similar tasks. To overcome these limitations, we focus on cross-embodied settings with shared robotic bases and heterogeneous end-effectors, and propose X-DiffVLA, a diffusion-based VLA model featuring a unified cross-embodied action head. X-DiffVLA can leverage the generative strengths of diffusion models to capture both the diversity and latent correlations in cross-embodied datasets. Specifically, we introduce Embodiment Forcing, a classifier-free guidance technique to implicitly steer action generation toward embodiment-specific functional components, capturing fine-grained structural nuances without explicit supervision. In addition, a Morphological Tree Diffusion approach is designed to strengthen behavioral correlations across diverse end-effectors, maximizing the transferability of heterogeneous demonstrations. Experimental results across RoboCasa and Isaac Gym, covering different embodiments from grippers to dexterous hands, show that X-DiffVLA achieves state-of-the-art performance, with improvements of 15.3% and 12.5%, respectively. Real-world evaluations further validate the robustness of the proposed framework and its effectiveness in scalable cross-embodied policy learning.
arXiv:2605.25063v2 Announce Type: replace Abstract: Scan-order design in laser directed energy deposition (LDED) is a delayed, path-dependent thermo-mechanical decision problem, because sequence quality becomes observable only after the complete deposition and cooling cycle. This work formulates LDED scan-order optimisation as a finite-horizon, permutation-constrained reinforcement-learning problem and develops a bilevel finite-element-teacher-labelled AI workflow. A surrogate-assisted teacher-guided optimisation loop learns the Abaqus-labelled response landscape and provides a tractable terminal-reward environment for policy training. A frozen Maskable Proximal Policy Optimization (MaskablePPO) policy is then used to generate legal scan-order candidates, which are independently validated through Abaqus thermo-mechanical simulations. The results show bounded, N-dependent policy-generation value rather than record-level dominance over the mature surrogate-assisted optimiser. The strongest scan orders are obtained by the teacher-guided surrogate loop, whereas PPO autonomously reaches competitive regions of the native response landscape, with stronger rank concentration at smaller track counts and a clear reliability boundary at longer horizons. The teacher-labelled landscape further supports a physically gated lexicographic reward hierarchy in which warpage admissibility is the primary constraint, plastic strain acts as a safety filter and residual-stress-related improvement is pursued conditionally within the admissible region. Validated sequences also reveal an interpretable scale-separated ordering tendency that combines global spatial dispersion with local structured grouping. This workflow provides a route from fixed scan-rule selection toward finite-element-teacher-validated policy generation, while preserving independent finite-element validation as the final physical gate.
arXiv:2605.25253v2 Announce Type: replace Abstract: Higher-dimensional automata (HDA) are a model of concurrency that models simultaneous execution of events using higher dimensional cells. HDA recognize languages of pomsets, a generalization of finite words whose letters are partially ordered. We prove a new algebraic characterization of HDA languages: a language of pomsets is regular if and only if it is the inverse image of a functor from the category of pomsets into a finite category. Furthermore, the language is definable in first-order logic exactly when it is recognized by an aperiodic category, generalizing the McNaughton-Papert theorem to HDA languages. We also investigate a notion of counter-free HDA, and show that if a language is accepted by a counter-free HDA, it must be definable in first-order logic. The converse, however, is still open.
arXiv:2605.26343v2 Announce Type: replace Abstract: Mechanistic interpretability seeks to explain a model's behaviour by finding its circuit: the sparse subgraph of the model's computation that is causally responsible for it. Automated methods have made this search systematic, but each one starts afresh for every behaviour, and the effort spent finding one circuit does nothing for the next. Circuit discovery has thus been automated, but not amortised. We ask whether circuit discovery can itself be learned. We frame it as a sequential decision problem over the computation graph of GPT-2 small, in which a policy removes edges until it reaches a compact subgraph that preserves the behaviour, guided by a faithfulness reward defined through causal intervention. A single policy trained across twelve behaviours recovers a faithful circuit for each, and once frozen it transfers to behaviours it never saw during training, recovering their known circuits without further search. A short warm-start improves these transferred circuits, returning far smaller ones than training from scratch. While the learned policy does not match a per-behaviour search on circuit size or cost, it shows that circuit discovery is a learnable, transferable procedure rather than a search repeated for every behaviour.
arXiv:2605.26755v3 Announce Type: replace Abstract: Multilingual fact verification requires evidence that is both relevant and sufficiently complete for reliable factuality prediction. However, existing systems often rely on search snippets, sentence-level evidence, or locally segmented passages, which can miss decisive context and produce fragmented evidence. To overcome these limitations, we propose SEEK, a Semantic Evidence Extraction with an adaptive chunKing framework that constructs coherent evidence chunks from full fact-checking articles by identifying semantic topic transitions and preserving local verification context. The constructed chunks are encoded using a multilingual encoder and then multilingual LLMs are finetuned using LoRA adapter for veracity prediction. Experiments on X-FACT and RU22Fact show that SEEK improves macro-f1 by up to 10% over semantic chunking, 19% over sentence chunking, and 20% over search-snippet baselines. Evidence completeness and significance analyses further show that SEEK preserves richer verification context and enables more reliable multilingual fact-checking.
arXiv:2605.26790v3 Announce Type: replace Abstract: Low-thrust trajectory design relies heavily on repeated evaluations of fuel consumption and transfer feasibility, which require expensive optimal control solutions. In this work, we show these quantities can be accurately approximated by machine learning surrogates, enabling fast and scalable evaluation across a wide range of scenarios. By increasing both dataset size and model capacity, we observe that low-thrust trajectory optimization follows a scaling law, with performance improving linearly with the logarithm of training data and network parameters, and no evidence of saturation within the explored regime. Guided by this observation, we construct a large-scale dataset using the proposed homotopy-ray strategy tailored to mission design requirements. A key is the introduction of a self-similar transformation, which allows generalization across semi-major axes, inclinations, and central bodies avoiding retraining. As a result, the same neural approximator can be applied to diverse orbital environments and mission classes. The proposed models accurately predict optimal fuel consumption and minimum transfer time for single- and multi-revolution transfers. Their performance and generalization are demonstrated on a public dataset, a multi-asteroid flyby problem from the Global Trajectory Optimization Competition, and an asteroid rendezvous mission design. The models and datasets are released as open-source to support the space community.
arXiv:2605.27792v2 Announce Type: replace Abstract: Atmospheric gusts expose wall-bounded turbulence to severe unsteady forcing, triggering complex non-equilibrium dynamics and extreme aerodynamic loads. In this study, direct numerical simulations are performed to investigate the spatiotemporal modulation of turbulent structures and the triggering mechanisms of near-wall extreme events under Gaussian-type transient forcing. The results reveal that high-amplitude gusts inject energy primarily into the streamwise velocity component, inducing a pronounced non-equilibrium phase lag during turbulent energy redistribution. This process produces hysteresis in wall friction and extends the relaxation time. Spectral and continuous wavelet analyses demonstrate that intense gust forcing suppresses high-frequency random fluctuations and reorganizes turbulent kinetic energy into low-frequency coherent structures. The characteristic frequency of these energetic structures locks onto the gust driving frequency, with a relative deviation of only $2.4\%$. Furthermore, the occurrence probability of extreme near-wall events, including extreme positive (EP) wall-shear-stress events and rare backflow (BF) events, increases by up to an order of magnitude under severe forcing. Using a two-step conditional averaging technique, we demonstrate that BF events are actively driven by intense, localized adverse pressure gradients and energetic ejections, which promote spanwise vortex roll-up in the buffer layer. By contrast, EP events are governed by energetic sweeps of high-speed fluid that compress intense spanwise vorticity into the immediate vicinity of the wall. These findings provide physical insights into non-equilibrium energy transfer and offer theoretical guidance for load alleviation and robust flow control of unmanned aerial vehicles operating in unsteady atmospheric environments.
arXiv:2605.28092v2 Announce Type: replace Abstract: Signal Temporal Logic (STL), has recently seen extensive development, owing to its rich expressivenes for autonomous planning and control. Nevertheless, existing verification and control synthesis methods are limited with respect to the complexity and degree of nesting of the formulae. In this work, we propose a novel approach to STL based on an operator acting on reachability value functions. This constitutes a new theoretical framework for handling complex multi-nested formulae while at the same time providing tools for on-line control synthesis. In contrast to focusing on the design of STL-based reachability (or control barrier) functions, we develop operator-based nesting rules directly. Our method's expressiveness is demonstrated both theoretically, where necessary and sufficient conditions for STL formula satisfaction are extracted, as well as in simulations with complex fragments.
arXiv:2605.28512v2 Announce Type: replace Abstract: Self-evolving scientific agents capable of conquering the hard tail of formal mathematics require Compositional Learning Behaviours (CLBs) -- the capacity to ground and recombine novel symbolic structures in context, beyond mere recombination of prelearned atoms. We propose S2B-LM, an adaptation of the CLB-evaluating Symbolic Behaviour Benchmark that removes numerical processing as a confound and adds chain-of-thought scaffolding to elicit rather than merely probe latent CLB competency. Cross-evaluating ten Lean~4 theorem provers on CLB competency in S2B-LM and miniF2F whole-proof performance, we find correlational and causal evidence of our claim: First, a necessary-condition analysis via quadrant test yields $p=0.004$, with model scale being ruled out as a confound. Second, extracting a CLB-encoding activation direction from DeepSeek-Prover-V2-7B using S2B-LM traces via Contrastive Activation Addition and applying it during miniF2F whole-proof generation on the AIME subset, CLB suppression collapses solve rate from $32.3\%$ to $2.9\%$, without loss of coherence, while suppressing a random activation direction of equal magnitude leaves it at $31.9\%$. Together, these results show that CLB competency is necessary but not sufficient for the hard tail of formal mathematical verification.
arXiv:2605.28863v2 Announce Type: replace Abstract: Imperfect-information multiplayer games test whether agents can act under hidden information, sparse rewards, and non-stationary opponents. We study these challenges in Big 2, a four-player imperfect-information card game. We develop a self-play RL framework for Big 2 that enables controlled comparisons between policy-gradient and value-approximating agents. Under a common environment, input representation, training budget, and evaluation protocol, PPO outperforms Monte Carlo Q approximation, SARSA, and Q-learning against random, greedy, and heuristic Big 2 opponents. We further find that moderate entropy regularization improves PPO by preventing the policy from becoming overly deterministic, and that current-policy self-play provides a stronger finite-budget curriculum than checkpoint self-play or fixed-opponent training. Together, these results show that Big 2 is a useful controlled setting for studying deep RL under imperfect information, multiplayer interaction, delayed rewards, and variable action sets.
arXiv:2605.29032v2 Announce Type: replace Abstract: Model-based reinforcement learning (MBRL) agents typically learn world models by minimizing predictive loss. However, powerful RL optimizers inevitably exploit minor model inaccuracies, leading to simulator exploitation and a reality gap where policies succeed in simulation but fail in the real world. We propose that the objective for learning simulators should be strategic robustness rather than predictive accuracy, and formulate this as a zero-sum minimax game between a model player and an adversarial policy player. We provide a comprehensive theoretical analysis: (1) an online learning guarantee showing the game is learnable with sublinear regret bounds; (2) a tractable critic-based simplification bounding the global policy-value gap by the local critic's loss; and (3) an Error-MDP duality, proving that finding the worst-case policy is formally dual to a standard RL problem where the reward is the one-step critic error. This duality yields a provably convergent active data selection algorithm. Experiments on continuous control tasks demonstrate that our approach reduces prediction error in strategically important regions by $1.5$-$2.2\times$ and enables policies trained purely in simulation to match near-optimal real-world performance.
arXiv:2605.29212v2 Announce Type: replace Abstract: Image quality in modern imaging systems emerges from the coupled effects of the sensor, optics, and computational reconstruction. Ultra-thin metalenses offer a path toward substantial miniaturization of optical modules, but practical designs often exhibit pronounced chromatic and field-dependent aberrations that necessitate computational reconstruction. In current metalens pipelines, reconstruction models are commonly trained and selected using distortion-based fidelity objectives, such as PSNR, yet these proxies can be weakly correlated with human preference and downstream utility, reflecting the well-known perception--distortion trade-off. We introduce MetaRanker, a human-in-the-loop active ranking framework that formalizes metalens image quality in terms of semantic interpretability, defined as the degree to which humans can reliably recognize objects and structures in the presence of optical artifacts. MetaRanker combines a probabilistic preference model with uncertainty-aware query selection, and leverages vision--language models to provide lightweight semantic priors. Importantly, these priors are used only to guide the sampling of informative comparisons; human judgments remain the primary supervision signal throughout. Across real-world and synthetic metalens datasets with distinct degradation profiles, MetaRanker produces rankings that align most closely with human assessments, while reducing the number of pairwise annotations required by approximately 80% relative to exhaustive pairwise evaluation. Finally, we show that standard image quality assessment metrics exhibit limited alignment with human interpretability in the metalens domain, positioning MetaRanker as a practical step toward perceptually grounded metalens evaluation and co-design.
arXiv:2602.22960v2 Announce Type: replace Abstract: World models based on video generation demonstrate remarkable potential for simulating interactive environments yet suffer from persistent difficulties in two key areas: maintaining long-term content consistency when scenes are revisited and enabling precise camera control from user-specified inputs. Existing methods based on explicit 3D reconstruction often compromise flexibility in unbounded scenarios and struggle to preserve fine-grained structures. Alternative methods rely directly on previously generated frames without establishing explicit spatial correspondence, thereby limiting controllability and consistency. To address these limitations, we present UCM, a novel framework for unified modeling of long-term memory and precise camera control via a time-aware positional encoding warping mechanism. To reduce computational overhead, we design an efficient dual-stream diffusion transformer for high-fidelity generation. Moreover, we introduce a scalable data curation strategy that utilizes point-cloud-based rendering to simulate scene revisiting, enabling training on over 500K monocular videos. Extensive experiments on real-world and synthetic benchmarks demonstrate that UCM significantly outperforms state-of-the-art methods on long-term scene consistency, while achieving precise camera controllability in high-fidelity video generation.
arXiv:2602.23135v2 Announce Type: replace Abstract: Edge classification on directed dynamic graphs requires modeling interactions between source and destination nodes exhibiting asymmetrical behavioral patterns and temporal dynamics. However, existing dynamic graph architectures largely rely on shared parameters for processing source and destination nodes, with limited or no systematic role-aware modeling. We propose DyGnROLE (Dynamic Graph Node-Role-Oriented Latent Encoding), a Transformer-based architecture that disentangles source and destination representations. By using separate embedding tables and role-semantic positional encodings, the model captures the distinct structural and temporal contexts unique to each role. Critical in limited-label settings, which are common in edge classification, is a self-supervised pretraining objective we introduce: Directional Role Alignment (DRA). DRA learns distinct but aligned source and destination embedding spaces by training source representations to retrieve their corresponding destination representations while a historical positive masking strategy excludes previously observed interactions from future negative comparisons. The masks introduce a temporally directional training signal in which node pairs progress monotonically from unseen to observed, after which the relationship is eligible only for further alignment. A comprehensive evaluation on four edge classification tasks across eight datasets demonstrates that DyGnROLE consistently outperforms a wide range of state-of-the-art baselines, highlighting the importance of role-aware representation learning and asymmetric pretraining for modeling complex directed interactions when labeled data is limited.
arXiv:2605.31483v3 Announce Type: replace Abstract: Despite Bengali being the sixth most spoken language in the world, no prior work has systematically evaluated hallucination in large language models (LLMs) for Bengali. We introduce BenHalluEval, a fine-grained hallucination evaluation framework for Bengali covering four tasks: Generative Question Answering (GQA), Bangla-English Code-Mixed QA, Summarization, and Reasoning. We construct 12,000 hallucinated candidates using GPT-5.4 across twelve task-specific hallucination types, drawn from three existing Bengali datasets, and evaluate seven LLMs spanning reasoning-oriented, multilingual, and Bengali-centric categories under a dual-track protocol that independently measures false-positive rate on ground-truth instances (Track A) and hallucination detection rate on hallucinated candidates (Track B). To jointly penalise both failure modes and prevent inflated scores from uniform response bias, we propose BenHalluScore, a dual-track calibration metric that ranges from 7.72% to 55.42% across models and tasks, revealing substantial variation in hallucination calibration. Chain-of-thought prompting, applied as a mitigation strategy, shifts response distributions without consistently improving hallucination discrimination. BenHalluEval establishes the first dedicated hallucination benchmark for Bengali and highlights the inadequacy of single-track and prompting-only evaluation approaches for low-resource language settings. The dataset and code are available at https://anonymous.4open.science/r/BanglaHalluEval-EB77.
arXiv:2605.31603v2 Announce Type: replace Abstract: Connector-based video unified models have demonstrated strong capability in instruction-grounded video synthesis, but integrating a large high-fidelity generator into the unified training loop is computationally prohibitive, limiting achievable visual quality. We therefore propose Lumos-Nexus, a training-efficient unified video generation framework that facilitates the development of strong reasoning-driven generation capabilities while significantly enhancing visual fidelity. Lumos-Nexus adopts a two-stage design: 1) During training, only a lightweight generator is aligned with the understanding block to learn to take in reasoning-driven semantic control. 2) During inference, we introduce Unified Progressive Frequency Bridging (UPFB) to progressively hand off generation to a high-capacity pretrained generator in the shared latent space, enabling coarse-to-fine refinement and producing high-fidelity videos without compromising reasoning quality. To fill the gap in reasoning-driven video generation benchmarks, we introduce VR-Bench, which assesses a model's capability to translate inferred intent into coherent and semantically aligned video content. Extensive experiments demonstrate that Lumos-Nexus achieves substantial gains in visual realism and temporal coherence on VBench, while exhibiting strong reasoning-based generative performance on VR-Bench. Code and models are available at https://jiazheng-xing.github.io/nexus-lumos-home/.
arXiv:2606.00097v3 Announce Type: replace Abstract: RocketSmith is an agentic system which intelligently automates the DFAM process for the development of high powered rockets suitable for launch. The system utilizes a large language model to orchestrate the execution of software tools to validate design characteristics such as flight stability and generate the parametric design components for the rocket assembly. A collection of subagents and skills enable optimization workflows of flight parameters via iteration in both zero-shot and human-in-the-loop workflows. With this system, four distinct high power rockets with various motor and assembly configurations were developed utilizing the unique design capabilities of additive manufacturing. These assembly components were fabricated using various FDM printers, manually evaluated for flight readiness, and flight tested at a launch event. From these tests, all rockets achieved a stable launch and two of the four rockets were successfully recovered in reflyable condition. The altimeter data validated that the rockets achieved an altitude 80% of the expected apogee predicted by the agentic system, establishing consistency between simulation and experimentation.
arXiv:2606.00305v2 Announce Type: replace Abstract: On-Policy Distillation (OPD) improves large language model reasoning by training a student model on trajectories sampled from its own policy under teacher supervision. Although OPD operates on trajectories, its learning signal remains token-level: it identifies deviations through high-loss tokens and repairs them through local reverse-KL correction. We show that this "trajectory-sampled but token-learned" mechanism cannot reliably bridge student trajectories toward teacher trajectories. About 30% of high-loss tokens fall into the low-divergence regime, indicating that many are surface-form mismatches rather than real reasoning forks. Moreover, even truly divergent tokens are difficult to repair with isolated token-level supervision, since reasoning failures often unfold as short-horizon distributional drift. We propose Trajectory-aware OPD (TOPD), which uses near-future trajectory information to identify real divergent states and distribute guidance across multiple future tokens. Experiments show that suppressing non-divergent high-loss tokens improves standard OPD from 47.8% to 48.2% average accuracy, while TOPD further improves performance to 52.2%, with gains on AIME24 from 60.0% to 63.3% and AIME25 from 46.7% to 53.3%.
arXiv:2606.00616v3 Announce Type: replace Abstract: Recent Vision-Language Models (VLMs) struggle with grounded reasoning, temporal consistency, and context aware planning in videos. We introduce pause-and-think-T, a reasoning-centric training dataset that encourages models to pause, reason over visual evidence, and produce concise, actionable responses. The dataset promotes structured reasoning prior to answer generation, guiding models toward human-like, scene-grounded assistance. We fine-tune a compact 4B-parameter model and evaluate it on our pause-and-think-B benchmark targeting contextual understanding and goal planning tasks. The model achieves 58.0% accuracy at 59x fewer parameters than Qwen3-VL-235B (58.9%), matching GPT-5.2 on scene understanding and surpassing GPT-4o. Beyond our benchmark, it also shows strong out-of-distribution performance on EgoThink and TempCompass, with substantial gains in affordance, assistance, attribution recognition, situated reasoning, and temporal order, without benchmark-specific training. Our results indicate that targeted reasoning supervision enables compact models to deliver actionable, visually grounded guidance while generalizing beyond training data, without requiring large-scale model expansion.
arXiv:2606.01215v2 Announce Type: replace Abstract: Current 3D spatial reasoning methods face a fundamental trade-off: neuro-symbolic 3D (NS3D) concept learners achieve interpretable reasoning through compositional programs but are constrained to closed-set concept vocabularies and simple programs; end-to-end 3D multi-modal LLMs (3D MLLMs) could handle complex natural language and open-vocabulary concepts but suffer from black-box reasoning without explicit spatial verification. We introduce APEIRIA, a neuro-symbolic 3D MLLM to bridge two paradigms by distilling symbolic reasoning patterns into MLLMs with natural language chain-of-thought. Our three-stage curriculum progressively builds reasoning capabilities: a) 3D perception alignment grounds object visual-geometric features to the LLM, b) CoT-SFT teaches query decomposition and stepwise verification from symbolic program traces, and c) CoT-RL extends reasoning patterns to open-set concepts and deeply nested instructions. By transferring reasoning patterns rather than concept-specific knowledge, APEIRIA preserves key NS3D virtues: transparent reasoning and modular interchangeability of planning and perception components. Evaluations on grounding, question answering, and captioning show that APEIRIA surpasses prior NS3D methods and matches state-of-the-art 3D MLLMs on 3D spatial reasoning datasets, unifying symbolic methods' systematic reasoning with MLLMs' flexibility. Code is available at https://github.com/oceanflowlab/APEIRIA.