Learned orchestration has emerged as a promising paradigm for large language model (LLM)-based multi-agent systems, in which a central orchestrator selects agents and controls the information they receive throughout a task. Despite rapid empirical progress, existing methods train the orchestrator by crediting each decision with its own advantage, implicitly assuming that the other decisions in the episode stay fixed. This assumption fails in orchestration, where the contribution of each decision to the final outcome depends on the decisions preceding it, so the gain of a joint update is not the sum of its per-decision effects. To address this challenge, we introduce \textbf{IAPO}, an interaction-aware policy optimization framework that explicitly accounts for interactions among decisions. Specifically, IAPO exactly decomposes the performance difference between two policies into the effects of individual decisions and their interactions, which we prove can account for half of the attainable gain, and then optimizes both from sampled trajectories while certifying that every update is an improvement. On long-input versions of GSM8K, MATH, DROP, MBPP and HumanEval, against fifteen published multi-agent and workflow-search systems on the same 7B executor, IAPO's deployed plans reach the strongest system's accuracy with fewer tokens, and removing the interaction term costs 7 to 17 points of accuracy on code.
In this article, we study deep nonparametric regression trained via stochastic gradient Langevin dynamics (SGLD), a variant of stochastic gradient descent that injects carefully calibrated Gaussian noise during optimization. Using a PAC-Bayesian analysis tailored to deep ReLU networks, we establish an excess-risk bound that simultaneously accounts for network approximation error, statistical complexity, and the algorithmic cost accumulated along the SGLD trajectory. Under a Bernstein condition and within a stable SGLD regime, the excess-risk bound contains a stochastic term whose coefficient is inversely proportional to the variance of injected noise, while the estimator retains the optimal rate for nonparametric regression under the H\"older smoothness condition up to logarithmic factors. This result provides an approximation-error-aware, rate-optimal statistical perspective on the widely observed generalization benefits of moderate SGLD noise. For per-sample clipped SGLD, we further establish an exact f-DP guarantee and the corresponding asymptotic $\mu$-GDP guarantee for the proposed estimator. We establish statistical convergence rates and differential privacy guarantees for DNN estimators trained using SGLD, while explicitly accounting for both network approximation error and the effect of injected noise in the excess-risk bound. Extensive simulation studies and analyses of two real datasets examine the finite-sample predictive performance and privacy-utility behavior of the proposed procedure.
Peer-review comments check what a paper does and judge why it matters, but no existing scheme distinguishes these two evaluative acts. We argue that making these acts explicit and measuring how reliably LLMs identify them is a prerequisite for responsible LLM-assisted review analysis: humans need to inspect review judgments and their grounds. This defines a more honest mission for NLP in peer review than full automation. We present ReviewMap, a framework that classifies each comment as assessment of claims and scholarly standards or calibration of contribution value and venue fit, with finer-grained attributes applied conditionally by act type. Applied to 500 ICLR~2026 papers through multi-model LLM consensus, the act layer proves reliably automatable under the operational guideline (inter-model $\kappa \geq 0.75$), supported by five-human validation on 470 development comments, while the attribute layer requires human oversight, defining a measured automation boundary. At the corpus level, 17% of reviewer-paper instances contain no panel-coded calibration, and the panel-coded calibration share is nearly $3\times$ as high in Strengths as in Weaknesses. This boundary informs three affordances: a review guideline that operationalizes the assessment-then-calibration workflow for reviewers, a per-comment rebuttal strategy scaffold for authors, and a discussion-phase dashboard for area chairs.
Predicting reaction impurities is a critical yet persistently data-limited challenge in pharmaceutical process chemistry. While machine learning models for reaction outcome prediction have advanced rapidly, benchmarking their ability to predict impurities remains impractical due to the absence of curated, publicly accessible datasets. We present OPRD-Imp, a dataset of 229 organic reactions drawn from 62 publications in Organic Process Research & Development, each annotated with experimentally observed impurities and their mechanistic origins. A total of 784 impurity entries is provided, classified into four origin categories. Reactions are represented as reaction SMILES and impurities as individual SMILES strings, with multi-step synthetic route connectivity preserved. All data were manually extracted from the source publications by trained chemists following a standardized extraction protocol, with expert chemical validation and manual structure drawing. OPRD-Imp is intended to serve as a benchmark resource for evaluating models that predict reaction impurities, providing experimentally reported impurity annotations and mechanistic origin labels that are not systematically represented in conventional reaction datasets.
Mapping behavioral intelligence in malware analysis reports to ATT&CK techniques supports threat analysis and defense, but reliable identification remains challenging. Supporting evidence is often scattered across long, heterogeneous reports, while closely related techniques share behavioral cues and differ only in subtle but decisive details. While large language models (LLMs) can use external knowledge to interpret reports, they may assign techniques based on plausible associations rather than sufficient evidence. We present MEvince, a knowledge-constrained LLM reasoning framework for report-level malware technique identification. Its core is a shared behavioral representation connecting report observations to technique definitions. This representation is built through agentic mining of technique documentation and supports two online stages: 1) primitive-guided extraction uses behavioral primitives to structure LLM-extracted observations, link them to source passages, and group dispersed evidence; 2) evidence-aligned retrieval matches this evidence to technique profiles through the same concepts, retaining likely candidates and confusable alternatives for LLM comparison of distinguishing behaviors. Controls at both stages validate evidence references and check technique-specific requirements. The representation also supports adaptation without parameter updates, where offline agentic audits diagnose concept weaknesses and refine extraction or retrieval policies. We evaluate MEvince against 11 baselines spanning four method families using a dataset of 610 public malware reports referenced by MITRE ATT&CK. MEvince improves exact-ID F1 over the strongest baseline by 17.4%, and its concept-guided retrieval increases gold-label candidate coverage from 31.6% to 47.9%. With a small adaptation set, offline refinement yields further gains and outperforms baselines that update model parameters.
Animating a reference image with a driver's motion requires a motion representation that captures face, hand, and body articulation but excludes the driver's appearance. Most body animation methods take motion from external pose estimators and render it with many denoising steps; compact learned motion representations mostly cover the face. We present FUHA, a real-time-capable feed-forward transformer that learns a compact motion embedding of the whole human and uses it for face, upper-body, and full-body animation. A Q-Former compresses the driver's features from a frozen visual encoder into this embedding. To disentangle motion from appearance, a novel contrastive loss pulls it toward the motion embedding of the driver's rendered keypoint image and away from that of the same person in another pose. At inference time, FUHA needs no external pose estimators, keypoints, or parametric body models. As a retrieval key, the motion embedding matches body pose better than the learned global embeddings of ten vision encoders. Its same-person retrieval rate is 0.1%, against 9.5% for its own DINOv3 backbone. On our upper-body benchmark, FUHA outperforms all baselines on every self-reenactment metric and on cross-reenactment identity similarity. It is competitive on face and full-body and runs at up to 38.9 FPS at 512x512 on a single RTX 4090.
When large language models do multi-step mathematics, a characteristic failure recurs: the model writes one wrong sign token mid-computation -- a cofactor sign, the minus of a recursive integration-by-parts step -- then executes the rest of the arithmetic correctly, and the answer comes out wrong. We ask where inside the network that sign is written. Across six open models spanning 14B to 675B parameters, a per-layer readout shows the wrong sign is committed late -- by MLP layers in the last quarter of the network (70-100% per model and task), a pattern present, where tested (Qwen), in the base model before instruction tuning. The mechanism varies: some models carry fixed "minus-stamp" layers that push a minus regardless of the right answer, others carry amplifiers of whichever sign was already chosen, and Qwen adds a downstream rescue layer that opposes the stamp. The circuit is causal: subtracting a single sign direction from the residual stream flips 13 of 14 held-out minus-written integration-by-parts errors on Llama-3.3-70B, with flips replicating on Qwen, Gemma, and Mistral-Small (24-70% per cell), and with representation-level specificity -- the same dose breaks correctly-minus answers while sparing correct-plus answers where the direction is surgical (0/72 on Llama), and matched random-direction and control-layer interventions are null. Core claims were preregistered with pass criteria frozen in git before held-out data was seen: 23 of 29 checks passed, all misses reported. But causality is not uniform: Phi-4 implements an equally well-localized circuit at the same late depth, yet the identical intervention flips almost nothing -- the sign is written redundantly across ~15 layers with margins several times larger (median 17.8 logits). Localization, we conclude, does not imply steerability: two models can share a circuit location yet differ enormously in how overturnable it is.
Tensor and sequence parallelism (TP+SP) and expert parallelism (EP) put collectives on the critical path of transformer layers. State-of-the-art fused kernels mitigate this with fixed roles or schedules, but this rigidity leaves large idle windows. Predicting a communication-computation split does not remove them either; a predetermined split cannot follow data readiness. We present KOSMOS, a library for self-scheduling persistent kernels whose workgroups use queues ordered by consumer demand to switch between communication and computation roles as data become ready, with no scheduler, per-call cost model, or tuned split. KOSMOS fuses every TP+SP collective of a transformer layer with its GEMM, and runs an entire expert layer, from router to fold, in one launch per direction. Evaluated on AMD Instinct MI350X and AMD Instinct MI355X GPUs, KOSMOS speeds up TP+SP end-to-end training in Megatron-LM by 1.10-1.28x in BF16 and 1.10-1.25x in MXFP8. KOSMOS accelerates end-to-end training of DeepSeek-V3 and Qwen3-235B proxies by 1.29-1.83x over DeepEP.
Kernel machines are an important class of models in machine learning. However, the question of how robust modern kernel machines are to adversarial examples has not been studied comprehensively. In this paper, we provide a thorough empirical investigation of the adversarial robustness of kernel-based models for both binary and multiclass classification tasks, via white-box and black-box evasion attacks. We evaluate the native robustness of kernel machines, random feature classifiers, and the relatively newer Recursive Feature Machines. In addition, we provide an algorithm for approximate adversarial training of kernel machines. Our experiments on classification tasks on image and tabular datasets strongly indicate that kernel machines are natively more robust than neural networks, i.e., their performance degrades more gracefully against adversarial perturbations.
In time-series forecasting, rollout errors are not merely one-step fitting discrepancies: once fed back into the evolving state, they can perturb future predictions. We formalize this effect through local rollout consistency, which compares the next prediction produced under teacher forcing with that produced under free running. In principle, this consistency can be enforced directly by tracking the residual-induced perturbation of the next state. In practice, however, the residual-to-state alignment operator is often difficult to design when the input state and output block have different dimensions, heterogeneous components, or architecture-dependent layouts. We therefore use a first-order local surrogate that penalizes the Jacobian-vector product of the predictor along the actual feedback perturbation, suppressing precisely the directions through which self-generated errors remain visible to future predictions. We further give the same loss a geometric interpretation near a predictive manifold: by controlling residual-induced directions that probe the normal bundle, the objective yields theoretical conditions under which free-running trajectories remain inside a stability tube around the manifold. Experiments on time series forecasting tasks support this view, showing that the proposed regularization improves predictive accuracy. More broadly, our results suggest a training principle for time-series forecasters and sequential predictors: a model should not only make small one-step errors, but also learn not to let its own errors echo into the future. Code is available at https://github.com/Wangzzzzzzzz/Echoes-Of-Error.
Long-horizon role-playing dialogue requires consistent character identity, reliable memory, and context-sensitive interaction. We present a bounded recursive self-improvement (RSI) framework that revises original-character memory, stable persona instructions, and a policy for constructing turn-level prompts. Repair outcomes also inform revisions to the updater's diagnostic rules, edit routing, and candidate selection. We specify the updater as a versioned policy with explicit inputs, editable fields, acceptance gates, and stopping conditions. An internal evaluation covers 50 original characters and five models, with two 100-round conversations per character-model pair: 500 conversations per condition. Across the pooled evaluated replies, the bad-response rate decreases from 14.00% (6,973/49,824) to 4.94% (2,459/49,793), a 9.06-percentage-point reduction and a 64.71% relative reduction. In 600 blinded human comparisons, RSI receives 294 wins, 187 ties, and 119 losses, yielding a 71.19% win rate among decisive judgments. The results support improved aggregate dialogue quality for the complete framework on the internal test distribution.
Language agents increasingly work for users on the open Web, and finance is where trusting them is hardest. The data must come from the Web, the strategy lives in the user's head, and correctness depends on time, so an agent can fail silently: a strategy the user never asked for, unverified data, or a backtest that uses the future yields a report as convincing as a correct one. Existing benchmarks close the world somewhere, fixing the task, the data, or the timeline, and thereby remove these failures instead of measuring them. We introduce FINCOSMOS, an open-world evaluation paradigm for agents doing financial work. A simulated user holds a hidden strategy sampled from a compositional grammar and reveals it only when asked. The agent receives no data, so it must acquire data from the Web, build its own backtest under a data cutoff, and deliver verifiable results. Grading needs no reference answer. Programs verify point-in-time correctness and recompute every reported metric, the delivered strategy is checked against the hidden specification, and a run is accepted only when the user is satisfied and none of 15 veto checks fails. Across four coding agents and a model without tools in 150 sessions, no session is accepted: every session that the simulated user signed off fails a program check, and agents raise only 54--64\% of the details they were not told. Human experts attribute 88.33\% of failures to the agents. With FINCOSMOS feedback, an agent with persistent memory learned all ten tasks.
LiDAR-Inertial-Visual Odometry (LIVO) integrates complementary LiDAR, camera, and IMU measurements to achieve robust state estimation. However, most existing methods assume a fixed relative pose and accurate time synchronization between sensors. Toward modular and heterogeneous robotic systems, this paper presents FREE-LIVO, a tightly coupled odometry method that requires neither a rigid LiDAR-camera rig nor LiDAR-camera time synchronization. The proposed method uses a pretrained image feature matcher to establish correspondences between a camera image and a dense LiDAR intensity image generated from a single LiDAR scan using a generative model. From the resulting 3D point-to-pixel correspondences, we construct reprojection factors that directly constrain LiDAR and camera poses at different acquisition times. These factors are jointly optimized in a single factor graph together with constraints from point cloud registration, visual feature tracking, and the IMU measurements associated with each sensor. On six sequences from the public M3DGR dataset, FREE-LIVO achieves an average absolute trajectory error (ATE) RMSE of 0.45 m without provided LiDAR-camera extrinsic parameters and exhibits minimal accuracy degradation under time synchronization errors of up to 0.5 s. Furthermore, experiments with the LiDAR and camera mounted on separate robots demonstrate that the method can jointly estimate the trajectories of independently moving sensors.
How can an agent handle a real-world task end to end? Frontier models with their own harnesses still fall short on real-world benchmarks: on Terminal-Bench 4.0, GPT-5.6 Sol with Codex resolves 37.3% of the tasks and Claude Opus 5 with Claude Code 51.8%. These tasks fail at the step from reaching a resource to applying it, which we call the last mile to the real world. Existing approaches put the resource within reach but leave the competence to apply it with the caller, so each harness rebuilds that competence. We introduce agent-native primitives, reusable units that equip a resource with a resident LLM that applies it in context and communicates with other primitives, and HEELO (Harness Engineering around Evolving Agent-Native Primitives for the Last Mile to Real-WOrld Tasks), a harness engineering framework that builds each task's harness from a library of such primitives. HEELO composes the primitives a task needs into a task-specific graph, verifies the result against evidence from the environment, revises the primitives a failure implicates, and admits new or improved primitives after validation. We instantiate HEELO with eight primitive families, each built for one kind of task, and evaluate them on more than twenty benchmarks. Dev-Primitives raise Terminal-Bench 4.0 resolution from 37.3% under Codex to 55.5% on the same backbone, Harness Primitives composed per task raise Terminal-Bench 2.0 Pass@1 from 60.0% for the seed harness to 72.2%, Tool Primitives raise ToolBench pass rate from 55.8% to 75.1%, and Navigation Primitives raise GAIA-Text-103 accuracy from 45.6% to 58.3%. On the 46 real-world tasks of our heelo.dev platform, Claude Opus 5.5 with Claude Code and GPT-6.1 Sol with Codex complete 9 each on their own and 39 and 35 when connected to HEELO through our MCP server, with 1.9$\times$ and 2.3$\times$ less time and 1.55$\times$ lower cost per completed task.
Test-time compute (TTC) improves language model responses by spending additional computation at inference time, and the right amount of compute depends on the prompt. Adaptive TTC methods allocate compute based on signals of response quality such as observed rewards, often by stopping once the signal exceeds a fixed threshold, and generally measure cost by the number of generations or tokens. But in online serving, latency is also a cost: it grows with server load, so a method using a fixed threshold spends too little compute when generation is fast and too much when it is slow. For Best-of-$N$ sampling, a widely used TTC method, we propose a two-stage adaptive algorithm that draws an initial batch, then uses the observed rewards and the batch's latency to decide whether to draw a second batch and at what size by comparing expected improvement against predicted cost: it draws one when the first batch's best reward is below a threshold that decreases with the first batch's latency, without explicitly measuring server load. We map each reward to its value under the cumulative distribution function (CDF) of the prompt's reward distribution, so the same tradeoff parameters can be reused across tasks; a regression fitted offline on other prompts predicts a distribution for the CDF value of the first batch's best reward, giving the expected improvement in closed form. We evaluate the method by simulation on a preference-based alignment benchmark, measuring generations-equivalent cost: generations plus latency, converted to generations. The method reaches Best-of-$N$'s reward at lower generations-equivalent cost, and conditioning on the measured first-batch latency saves at least as much of that cost as fixed-threshold (latency-unaware) baselines, and more when latency is highly variable, when it grows with batch size, and when the second batch's latency is highly predictable from the first's.
Accurate forecasting of renewable generation requires not only point predictions but also reliable estimates of the uncertainty attached to them, particularly at extended horizons. Machine learning models are increasingly used to predict solar output, but they typically report a single value and capture the associated uncertainty poorly. The standard remedy is to train a model to predict the parameters of an assumed probability distribution directly. Here we develop an alternative two-stage method, motivated by synthetic solar signal generation, in which the forecast error is modeled separately from the expected irradiance. A first long short-term memory (LSTM) network predicts the expected global horizontal irradiance, and a second network models the distribution of the forecast error, which is added back to the mean to give both the forecast and its uncertainty. Using historical weather data from the National Solar Radiation Database and a probability integral transform evaluation, we show that explicitly modeling the error brings the predicted standard deviation markedly closer to the historical value, reducing the deviation between the two by 34.6 percent for a standalone LSTM and by 56.5 percent when the LSTM is coupled to Gaussian process regression, the most reliable of the configurations tested. Because the method models the total forecast error directly, it captures the full predictive uncertainty that matters for decisions, rather than only the model contribution that dropout-based approaches estimate.
Projector-based adversarial attack aims to physically manipulate real-world scenes by projecting adversarial patterns, thereby causing deep image classifiers to produce incorrect predictions. However, existing stealthy projector-based adversarial attack methods model the project-and-capture process in a single and static view, limiting their view-invariant capability and practical applicability in dynamic environments. In this paper, we introduce View-Invariant Projector-Based Adversarial Attack (VIPA), a novel method designed to overcome these limitations by achieving both view-invariant and classifier-agnostic adversarial attacks in the physical world. We first leverage a view-invariant projector-camera system simulation method to model the physical interactions between projected patterns and real-world surfaces. To ensure robustness and stealthiness of the attack across different views, VIPA optimizes adversarial projections by aggregating simulated attack losses from multiple views. This joint optimization across diverse views helps maintain robustness regardless of the viewpoints. Finally, the optimized patterns are physically projected into real-world scenes, where they successfully fool various classifiers from different viewpoints, thereby enabling robust and practical view-invariant adversarial attacks. Our experiments in both targeted and untargeted attacks demonstrate that VIPA consistently achieves higher attack success rates than existing methods, while also enhancing stealthiness and ensuring minimal perceptual degradation across all tested views.
We present IMUDINO, the first foundation model (FM) for inertial sensing to achieve competitive performance across diverse sensing tasks including human activity recognition (HAR), health, motion capture, and odometry. Prior inertial FMs often focus on one subfield of inertial sensing and sometimes underperform state-of-the-art supervised methods (e.g., in HAR), making their scope considerably narrower than FMs for vision and natural language. Moreover, they often require fixed input length and sampling rate, and generalize poorly to unseen sensor locations. To build a general-purpose foundation model for inertial sensing, we adapt the DINO self-supervised learning (SSL) paradigm to inertial data through a set of semantics-preserving motion transformations that simulate variations in motion pattern and sensor setup. For model architecture, we use a plain transformer encoder with frequency-aware positional encoding that natively supports varying input length and sampling rate. This simple setup proved remarkably effective for learning global and dense features generalizable across users, sensor setups, tasks, and, in the case of robot odometry, embodiments. We hope IMUDINO can serve as a backbone for inertial sensing and kickstart the development of general-purpose inertial FMs.
Medical images contain topological patterns, such as connected components and loops, at scales ranging from small lesions to organ boundaries. Although hierarchical Vision Transformers have advanced medical image analysis, they do not explicitly incorporate topology throughout their multiscale representations. We propose Topology-Reinforced Swin Transformer (TRiST), a framework that integrates topological information into feature construction, refinement, and attention. First, a Hierarchical Multi-Scale Topology Representation algorithm computes 2D persistent homology (H0 and H1) independently at four native Swin stages, producing descriptors aligned with each stage's token grid rather than downsampling fine-scale topology. Second, a Band-Aware H0/H1 Residual Refinement module uses gated residual networks to process short-, medium-, and long-persistence information separately. Third, a Persistence-Band Progressive Attention Bias converts stage-specific descriptors into learned key-saliency scores that directly modify shifted-window attention logits. We instantiate TRiST as Topo-SwinV2-B for classification and Topo-Swin-UNet for segmentation, preserving the underlying Swin-based backbones. Experiments on five medical image datasets show consistent improvements over Swin-based baselines and competitive performance relative to existing topology-augmented methods.
Real-world environments are messy, with paths often blocked by everyday objects like chairs and toys, requiring robots to perform interactive navigation by manipulating obstacles to clear a path. Existing methods rely on extensive simulation-based learning or per-scene instructions specifying which objects to avoid or manipulate. They struggle to generalize across novel objects, diverse environments, and varied contexts such as a cart of fragile glassware that should be avoided while an empty one can be pushed. We formalize this problem as open-world interactive navigation and present MessyNav, a zero-shot framework based on vision-language models (VLMs). At its core, MessyNav performs explicit VLM-based affordance reasoning before task planning to decide whether to swipe, pick up, or avoid an object and produce a high-level plan. Evaluated on a mobile manipulator across real-world indoor and outdoor environments, MessyNav improves success rate by 27 percentage points (pp) over the strongest baseline in real-robot trials and achieves statistically significant gains in a larger-scale offline evaluation, without task-specific training.