arXiv AI

arXiv:2607.11250v2 Announce Type: replace-cross Abstract: Exploration is essential for reliable autonomy in multi-agent systems, yet it remains unclear whether large language model (LLM) agents can explore effectively when interacting with one another. We show that modern LLM agents often fail to do so, exhibiting myopic and polarized interaction patterns that lead to suboptimal coordination and increased regret. We formalize this challenge as the Multi-Agent Exploration problem, modeling it as a partially observable stochastic game (POSG), in which agents must probe peers to infer their capabilities and identify effective interaction strategies. To address this, we introduce Multi-Agent Contextual Exploration (MACE), a lightweight framework that explicitly promotes exploration through structured peer selection. Across both contextual and parametric diversity settings, exploration behavior and downstream task performance substantially improve when MACE is applied. We further provide theoretical insight that MACE achieves a sublinear expected-regret bound. Overall, our results highlight a fundamental limitation of current LLM agents and underscore the importance of explicitly guided exploration for reliable multi-agent autonomy.

arXiv:2607.05475v2 Announce Type: replace-cross Abstract: Deploying Large Language Models (LLMs) on mobile devices enhances privacy and reduces latency, but is severely constrained by hardware inefficiencies. We present the first comprehensive, cross-layer measurement study of mobile LLM inference, spanning five mainstream frameworks (e.g., llama.cpp and GENIE) and three hardware backends (CPU, GPU, and NPU). To enable this analysis, we develop PowerBench, a fine-grained profiling tool that measures LLM throughput and provides the first backend-specific energy attribution for mobile LLM inference, going beyond traditional device-level measurements. We further conduct hardware microbenchmarks to characterize the memory-system performance and compute capabilities of mobile CPUs, GPUs, and NPUs. Our study yields three critical insights: (1) Framework-induced performance gaps are substantially amplified on NPUs, reaching up to $10\times$ due to differences in offloading strategies, hardware-underutilizing operators, and activation quantization strategies. (2) We identify a distinct ``phase split,'' in which NPUs excel at compute-bound prefilling, while CPUs outperform all other backends in memory-bound decoding. This split is driven by NPUs' preference for large, fixed-shape workloads, which conflicts with the small-kernel, dynamic nature of decoding. (3) Backend-specific profiling uncovers substantial scheduling headroom missed by prior work. Suboptimal thread configurations, uncoordinated NPU sleep latencies, and CPU polling intervals result in unnecessary energy consumption. For example, tuning CPU frequency alone reduces energy consumption by up to $39.7\%$. Leveraging these findings, we estimate that an energy-oriented configuration can reduce energy consumption by up to $53.2\%$ on the NPU backend across three datasets, highlighting the substantial potential for further efficiency improvements in mobile LLM deployment.

arXiv:2607.04593v2 Announce Type: replace-cross Abstract: Vision-Language Models (VLMs) have demonstrated impressive capabilities across different tasks, but their computational cost is dominated by the large number of visual tokens fed to the language model. Existing token reduction methods rely on attention-based scores or pairwise similarity, without an explicit semantic representation of each token. We introduce TORINO (TOken Reduction via Interpretable coNcept Overlap), a plug-and-play framework for adaptive visual token reduction in VLMs that requires no fine-tuning of the underlying model. TORINO leverages Sparse Autoencoders (SAEs) to project visual tokens into an interpretable latent space where token relationships can be analyzed through shared concept activations. Specifically, we define concept overlap as the degree of agreement between active SAE latents and use it to group tokens that share semantic content. Reduction within each group is then performed by either pruning or merging, providing a unified framework that preserves semantically important visual information while removing redundancy. Unlike fixed-budget approaches, TORINO dynamically adapts the reduction rate to input complexity, allowing different images to retain different numbers of tokens. Experiments across multiple vision-language benchmarks show that TORINO achieves favorable efficiency-accuracy trade-offs, reducing the number of visual tokens with minimal performance loss. Source code released at https://eidoslab.github.io/TORINO.

arXiv:2606.31148v3 Announce Type: replace-cross Abstract: 3D Visual Grounding (3DVG) aims to localize target objects in 3D scenes given natural language descriptions. Existing approaches typically perform reasoning over the entire scene, leading to ambiguous predictions and high computational cost, especially in cluttered environments. We observe that many referential expressions rely on local spatial context and often correspond to restricted spatial regions rather than the full scene. Motivated by this insight, we propose PruneGround, an effective plug-and-play framework for 3DVG built upon three key components. First, we introduce Language-Guided Spatial Pruning (LGSP), which leverages a frozen Vision Language Model (VLM) to identify language-relevant regions, thereby reducing spatial computation and grounding candidates in the narrower search space. Second, we propose MultiView-Conditioned Description Reformulation (MCDR), which decomposes complex expressions into simplified target--anchor relations and augments missing spatial cues through multi-view reasoning. Finally, we propose LLM-Grounder, which repurposes a detection-pretrained spatial LLM into a language-conditioned grounding model by aligning point cloud and linguistic representations within the pruned region. Extensive experiments on the three most popular point cloud benchmarks demonstrate that our method achieves state-of-the-art results on all three ScanRefer settings and on 9 out of 10 Nr3D/Sr3D settings. Code and models are publicly available: https://github.com/leduckhai/PruneGround

arXiv:2606.22995v2 Announce Type: replace-cross Abstract: Group-based Reinforcement Learning (RL) has significantly enhanced Large Language Models (LLMs) in agentic scenarios. To achieve finer-grained policy updates, recent agentic RL frameworks have shifted from trajectory-level to step-level training. However, long-horizon agentic RL suffers from severe reward sparsity and delay, as feedback is often deferred for dozens of interaction steps. While existing step-level frameworks refine training granularity, their credit assignment remains coarse-grained and still treats agent exploration as isolated, linear trajectories. This oversimplified perspective ignores the inherent graph structure of state transitions, leading to high-variance state-value estimation and myopic, localized credit assignment. To overcome these critical bottlenecks, we propose Group-Graph Policy Optimization (G2PO), a novel group-based RL algorithm tailored for multi-turn agentic tasks. G2PO explicitly transforms linear interaction trajectories into a global state-transition graph. By aggregating identical observations across different trajectories, we introduce group-aggregation state-value estimation that reduces sampling variance and trajectory-dependent bias. Furthermore, we redefine agent actions as transitions between state nodes and propose an edge-centric advantage estimation strategy. By globally standardizing Temporal Difference (TD) errors across the entire graph, G2PO explicitly identifies and prioritizes critical transitions that drive absolute task progress. Extensive experiments on representative long-horizon benchmarks-WebShop, ALFWorld, and AppWorld-demonstrate that G2PO substantially outperforms state-of-the-art prompt-based and RL baselines, achieving remarkable success rate improvements of up to 22.2% over GRPO.

arXiv:2606.18986v2 Announce Type: replace-cross Abstract: Recent advances in large language models (LLMs) have given rise to time-series question answering (TSQA), which formulates time-series analysis as natural-language question answering. However, directly feeding raw numerical series into LLMs suffers from a tokenization bottleneck: Byte Pair Encoding fragments continuous values into unstable tokens whose embeddings lack meaningful metric structure, resulting in the loss of magnitude, scale, and trend information. Prior methods use patch-based encoders that split the series into fixed windows, locking in one granularity that breaks patterns and hides exact timesteps, through a separate module that rarely transfers across datasets with different lengths or sampling rates. To address this challenge, we propose CADE (Contrastive Alignment with Direct Embedding), a novel framework for TSQA built upon two key components: direct timestep embedding and semantic alignment. The proposed framework maps each timestep directly into the LLM embedding space through a point-wise linear encoder and MLP projector, preserving exact index-level access while eliminating the need for patching and padding. To further bridge the semantic gap between time-series and language representations, we introduce a novel one-directional supervised contrastive loss that aligns time-series embeddings with frozen class-name text anchors. Experimental results on the public Time-MQA benchmark demonstrate that our framework consistently improves performance across six TSQA tasks, outperforming both open-source and proprietary LLM baselines.

arXiv:2606.13535v5 Announce Type: replace-cross Abstract: Particle physics collider experiments provide Rivet routines as part of the analysis preservation strategy for model-independent measurements. Rivet is a C++ toolkit that allow new theoretical models to be compared to the measurements, thus aiding the development and tuning of Monte Carlo event generators as well as searches for physics beyond the Standard Model. However, analysis coverage is known to be incomplete, with only 39% of measurements having documented and publicly available Rivet routines. In this article, we design and implement an automated workflow based on Large Language Models with the goal of providing the missing routines. This multi-step workflow, referred to as AgentRivet, extracts the physics analysis information from published papers and writes the missing Rivet routines, with intermediate code- and physics- reviews as part of an autonomous quality control. We report the results obtained using commercial Large Language Models, provided by OpenAI, Anthropic, and Google, for two recent measurements from the ATLAS and CMS experiments. We find that AgentRivet produces competent Rivet routines with few syntax errors. The physics fidelity of the routines is reasonable and follows the explanations given in the relevant publications. Nevertheless, physics-implementation issues do arise and are investigated using the artefacts produced by AgentRivet. The majority of physics implementation issues arise from subtle-but-ambiguous definitions in the given publication, although some models struggle to implement complex observables even when clear definitions are given.

arXiv:2606.10660v2 Announce Type: replace-cross Abstract: Purchased artificial intelligence (AI) inference belongs in Scope 3 Category 1 of a corporate greenhouse gas inventory, yet practice omits it or applies a spend-based factor that gives 1.7 to 12.3 times the physical estimate and cannot tell one model or region from another. This paper sets out a token-based method and open dataset: a cradle-to-gate boundary, four tiers of activity data, an effective-token basis, a four-term factor (serving electricity, embodied hardware, data-centre construction, disclosed final training runs) and a rule resolving any model identifier and region to a factor. Inference footprints span twelve times between model classes and nine between regions. The mid-class factor, 0.039 kg of carbon dioxide equivalent (CO2e) per million billed tokens at the world-average grid, is the low end of a stated structural range: a less efficient serving configuration raises it 2.8 times, and replacing the training term with all model development, on its own boundary, gives 1.5 to 2.9 times it. Input bands alone give -18% to +29%. Published per-query energy estimates that can describe a mid-class model are 2.4 to 7.2 times the mid-class energy, above its structural range in four of five rows; those that can describe a frontier-class model bracket it at 0.7 to 1.6 times. Worked examples give 8.5 kg CO2e a year for a 200-person firm, and 2.1 tonnes for fifty developers using agentic coding tools on a mid-class model or 8.3 tonnes on a frontier-class one. The contribution is a traceable figure; accuracy against metered energy is untested.

arXiv:2606.08408v2 Announce Type: replace-cross Abstract: We extend activation steering to diffusion language models (DLMs) and study a novel problem that arose due to the inference mechanism of DLMs: Modifying a text in-place to manifest a different concept. We propose TimpaTeks, an automatic in-place text modification mechanism using DLMs. Experiments on IMDB movie reviews (sentiment) and a synthetic Cats and Dogs Dataset (arbitrary, more unconventional concept steering) show that TimpaTeks provides a feasible novel mechanism to steer diffusion language model outputs in-place. TimpaTeks enables in-place modification while simultaneously lowers sentence perplexity and retaining the original sentence structre without the need of instruction tuned models. TimpaTeks is also computationally cheaper than prompt-based DLM steering, as it performs denoising in-place rather than constructing an additional prompt-conditioned output sequence.

arXiv:2606.03602v2 Announce Type: replace-cross Abstract: Causal discovery from observational data remains challenging due to the fundamental limitations of purely statistical methods, such as statistical distinguishability within equivalence classes and sensitivity to finite sample sizes. While large language models (LLMs) offer a promising source of domain knowledge to complement statistical inference, existing LLM-augmented methods are vulnerable to LLM errors and incur high token costs. Moreover, reliance on a single data-centric algorithm can make results sensitive to algorithm-specific biases. To address these limitations, we propose CauTion, a framework that reliably integrates LLM domain knowledge into an ensemble of statistical causal discovery algorithms through consensus filtering and LLM reliability estimation. CauTion proceeds in three stages. First, an algorithm ensemble utilizes a consensus voting to resolve up to 96% of edges on which algorithms agree, achieving near-perfect accuracy on the filtered consensus edges. Second, a trust-calibrated arbitration mechanism estimates the relative reliability of the LLM and the algorithms via an annotation-free trust calibration procedure, which is then utilized to govern a trust-weighted voting process that restricts LLM arbitration exclusively to edges with unreliable algorithmic evidence. Third, a cycle repair step is applied to guarantee the final causal graph is validly acyclic. Experiments on six datasets demonstrate that CauTion consistently outperforms both data-centric and LLM-augmented baselines, with larger gains on larger graphs and strong robustness to LLM errors. Code is available at https://github.com/OpenCausaLab/CauTion.

arXiv:2605.29468v2 Announce Type: replace-cross Abstract: Large language models (LLMs) are increasingly used to support scientific work, but it is unclear whether they uphold responsible conduct of research (RCR) norms or help undermine them. We introduce SciIntBench, an adversarial benchmark of 1350 prompts across 270 scenarios, ten RCR categories, and three scientific domains. Each scenario appears in five gradations, from Most Overt to Most Covert adversarial prompts, and an additional Benign version, allowing us to jointly measure framing-sensitive refusal of misconduct and helpfulness on legitimate requests. We evaluate 16 commercial and open-weight LLMs from six providers (2024--2026), producing 21,600 responses. We assess the responses using LLM judgments validated by human evaluation, achieving high agreement ($\kappa$ = 0.857). We find that, under the most covert framing condition, models fail to refuse requests for research misconduct in 71.7\% of cases, on average. Refusals vary by RCR category, with weaker boundaries around transparency, plagiarism, and fabrication.

arXiv:2605.29001v2 Announce Type: replace-cross Abstract: A paraphrase-quality audit of MathCheck (ICLR 2025) detected 4 semantically incorrect paraphrases in 129 groups (3.1%); removing them drops GPT-4o from rank 2 to rank 4 and elevates Claude Haiku and DeepSeek V3 above it. Cross-model unanimity found these errors automatically for under $10; in our own dataset the same protocol found that 47% of auto-generated connective-variation paraphrases were semantically incorrect. That flaw compounds a deeper measurement gap. On the theorems every model answers correctly in canonical form, a semantically equivalent rewording still flips the answer for 5-42% of them: Claude Haiku 4.5 flips a known-correct answer about 9x as often as DeepSeek V3 (41.9% vs. 4.7% on the 43 theorems all models share), a difficulty-controlled gap we call the Conditional Inconsistency Rate (CIR). The CIR ranking replicates on an independent 103-theorem set. Raw Semantic Consistency Rates (SCR) span 50-82% across 9 models, but this spread is mostly a mechanical consequence of compounding accuracy over k ~ 6 paraphrases; what CIR isolates is the within-theorem clustering of failures. Because no model dominates on all paraphrase families, the families a benchmark includes implicitly fix its model ranking (a No-Free-Benchmark corollary). FormInv supplies the audit protocol, CIR and SCR as invariance measures, evaluated on 9 models across 366-811 items whose canonical statements are Lean4-verified, and FormInvSelector for regime-aware model selection.

arXiv:2605.23916v3 Announce Type: replace-cross Abstract: AI agents often pick tools from registries, where each provider writes its tool's description; we ask whether sales language there changes which tool an agent picks. We built pairs of listings that differ in one way (praise, a verifiable specification, or list order) and asked two OpenAI models to call one tool. In a preregistered study, stacked praise (four kinds) raised a tool's pick rate by about 43 percentage points, matching or beating a verifiable specification. Praise also pulled some picks toward tools whose numeric limit was too small. With identical listings, the first-listed tool was picked about 72 points more often. Structured fields helped agents pick the capable tool; sales text beside them reduced or erased that gain. An exploratory study of five models shows the same direction; a "Sponsored" label dampened it, and a system-prompt warning moved one model. Registries could list limits as fields, hide sales text from agents, and randomize order. Stacked praise replicated on held-out domains; no single kind did after correcting for multiple tests. Preregistered results are provisional until blind phrase ratings are complete, and cover two small models.

arXiv:2605.22940v2 Announce Type: replace-cross Abstract: Entropy reduction is not a safety criterion: it can leave predictions unchanged or accompany increasing risk for a protected group. We study entropy-regulated learning under separately specified task, risk, and intervention constraints. A noise-calibrated log-determinant functional yields an exact spectral force formula and a metric decomposition into observable and prediction-null velocities. For scalar feedback, we characterize the complete interval of coefficients preserving a task-descent fraction and meeting an entropy-rate request; restricting intervention to observable directions can make that interval empty. For several protected risks, we derive an exact minimax certificate of feasibility within a specified subspace and norm budget, an exact robust extension for uncertain constraint rows, and a cone-projection formula for the entropy reduction remaining after protection. These formulas use standard convex geometry but identify when apparently active entropy control has no admissible effect. Barrier comparisons, finite-step curvature and gradient-error margins, and fresh-data audits give distinct conditional guarantees for risk thresholds. Human authority enters through the specification and release policy; the mathematics does not identify compression with alignment or certify human well-being. A memorization counterexample separates fresh-input information from sample-level information, and a noisy sample-code assumption supports a conditional generalization bound. Reproducible mathematical checks and a small real-data experiment expose both feasible interventions and failures. The resulting theory characterizes admissible entropy regulation and its limitations, rather than asserting universal accuracy or safety improvements.

arXiv:2605.15908v2 Announce Type: replace-cross Abstract: Natural images are continuous, yet most generative models synthesize them on discrete grids, limiting resolution-flexible generation. Continuous neural fields enable resolution-free rendering, but prior methods introduce continuity only at the decoding stage as an interpolation module, leaving the generative latent space discretized and reconstruction-oriented. We propose RaPD (Resolution-agnostic Pixel Diffusion), which performs diffusion in a continuous Neural Image Field (NIF) latent space. RaPD bridges this reconstruction generation gap with Semantic Representation Guidance for generation-aware latent learning and a Coordinate-Queried Attention Renderer for coordinate-conditioned, scale-aware rendering. A single denoised latent can be rendered at arbitrary resolutions by changing only the query coordinates, keeping diffusion cost fixed. Experiments demonstrate superior generation quality and resolution scalability.

arXiv:2605.12153v3 Announce Type: replace-cross Abstract: We present the Curated Industrial Developer Repository (CIDR), a large-scale dataset of real-world software repositories collected from industrial partners. The dataset comprises 4,687 repositories spanning 84 programming languages, totaling 1,033 million raw lines of code (670,3 million logical lines), along with structured metadata at the repository level, full version control history, and engineering-practice attributes such as continuous integration usage and the presence of automated tests. All repositories were collected, filtered, and anonymized through a multi-stage pipeline developed specifically for this purpose. We additionally report an exploratory fine-tuning study that adapts a 3-billion-parameter code language model to CIDR and quantifies the effect on held-out enterprise code. CIDR is intended to support research in code intelligence, software quality analysis, developer tooling, and related software engineering tasks. Access to CIDR is provided under a restricted license; details on eligibility and terms are available at https://fermatix.ai/#Contact.

arXiv:2605.08625v2 Announce Type: replace-cross Abstract: Time Series Foundation Models (TSFMs) excel at numerical forecasting but operate as black boxes lacking qualitative reasoning. Conversely, applying LLMs directly to temporal data introduces a modality gap: text tokenizers fragment continuous numerical values, degrading mathematical relationships and exploding sequence lengths, leading to computational overhead. To resolve this, we introduce STRIDE (Strategic Time-series Reasoning Injected via Distilled Embeddings), a novel framework natively integrating LLM reasoning into the continuous embedding space of TSFMs. Instead of discrete tokens, STRIDE distills reasoning traces into a lightweight LLM and dynamically projects its pooled hidden states as a continuous cross-modal prior into the numerical encoder. Jointly optimized via cross-entropy and quantile losses, STRIDE readily attaches to diverse architectures (e.g., Chronos-2, Timer-S1). STRIDE ranks #1 on the GIFT-Eval leaderboard (0.625 MASE, 0.423 CRPS) when ensembled with Synapse. On TFRBench, STRIDE consistently outperforms baselines in both forecasting accuracy and reasoning quality, with ablations confirming gains stem from strategic reasoning. Thus, injecting semantic reasoning as a continuous prior equips TSFMs with human-interpretable reasoning while improving predictive accuracy.

arXiv:2605.06557v2 Announce Type: replace-cross Abstract: Cooperative multi-agent reinforcement learning (MARL) benchmarks commonly emphasize aggregate outcomes such as return, success rate, or completion time. While essential, these metrics can obscure how agents coordinate and fail to distinguish policies that achieve similar outcomes through different interaction processes. We propose a process-aware evaluation framework that complements aggregate performance with diagnostics of interference, coordinated task completion, participation, and coordination efficiency. We evaluate the framework in STAT, a controlled spatial task-allocation testbed, and examine its broader applicability in two established cooperative MARL benchmarks. Across diverse value-based and actor-critic methods, we find policies with nearly identical return can exhibit substantial differences in coordination behavior. Controlled scaling in STAT reveals distinct process signatures for spatial, task, and agent scaling, while a reward-shaping ablation shows that explicitly penalizing assignment conflicts can alter coordination behavior with little change in task return. Results in the established benchmarks further show that the framework remains informative under different coordination mechanisms. These findings support process-aware evaluation as a complement to outcome-based benchmarking. The framework provides a common structure for selecting and interpreting environment-specific diagnostics that characterize not only what cooperative policies achieve, but how they coordinate to achieve it.

arXiv:2605.05524v2 Announce Type: replace-cross Abstract: Causal representation learning (CRL) seeks to recover latent variables with identifiability guarantees, typically up to permutation and component-wise reparameterization under appropriate assumptions. However, identifiability does not imply interpretability: latent semantics are typically assigned post hoc by alignment with known ground-truth factors. This limitation is particularly acute in scientific time series, where underlying mechanisms are unknown and discovering interpretable structure is a primary goal. In contrast, scientific observations (such as residue-pair distances, climate indices, or process sensors) are inherently semantic, as they correspond to named physical quantities. This raises a key question: can the interpretability of observations be transferred to the identifiable latent space? We propose MOSAIC (Module discovery via Sparse Additive Identifiable Causal learning), a sparse temporal VAE that integrates temporal CRL identifiability with support recovery over observed variables. MOSAIC identifies latent variables via regime-conditioned temporal variation, and recovers for each latent a sparse set of associated observations through an additive decoder, yielding module-level interpretability. We show that ANOVA main-effect supports are identifiable under general smooth mixing functions, with finite-sample recovery guarantees for a group-lasso estimator and empirical evaluation of MOSAIC's entropy regularizer. Empirically, MOSAIC recovers domain-consistent variable groups across RNA molecular dynamics, solar wind, ENSO climate, the Tennessee Eastman process, and a synthetic tokamak benchmark, enabling interpretable discovery of latent mechanisms in scientific time series.

arXiv:2605.04702v2 Announce Type: replace-cross Abstract: Identity-preserving text-to-video generation (IPT2V) empowers users to produce diverse and imaginative videos with consistent human facial identity. Despite recent progress, existing methods often suffer from significant identity distortion under large facial pose variations or facial occlusions. In this paper, we propose FaithfulFaces, a pose-faithful facial identity preservation learning framework to improve IPT2V in complex dynamic scenes. The key of FaithfulFaces is a pose-shared identity aligner that refines and aligns facial poses across distinct views via a pose-shared dictionary and a pose variation-identity invariance constraint. By mapping single-view inputs into a global facial pose representation with explicit Euler angle embeddings, FaithfulFaces provides a pose-faithful facial prior that guides generative foundations toward robust identity-preserving generation. In particular, we develop a specialized pipeline to curate a high-quality video dataset featuring substantial facial pose diversity. Extensive experiments demonstrate that FaithfulFaces achieves state-of-the-art performance, maintaining superior identity consistency and structural clarity even as pose changes and occlusions occur. Project Page: https://mdswyz.github.io/FaithfulFaces/