让下一篇重要论文，不再错过。

Data attribution methods for Language Models (LMs) are typically evaluated using downstream tasks like fact retrieval. However, existing evaluations vary multiple dataset properties simultaneously, making it impossible to isolate exactly why an attribution algorithm fails. To diagnose these failures, we explore the design space of LLM data attribution, with a primary focus on evaluation. We introduce the Controlled Attribution Benchmark (CATT) for fact tracing, in which we fine-tune LLMs on entirely fictitious knowledge, providing unambiguous ground truth for attribution. This synthetic setup allows us to systematically alter one dataset configuration at a time, such as document redundancy and lexical distractors. With this configurable evaluation, we demonstrate that redundancy and distractors consistently degrade attribution performance. Rather than proposing a new estimator, we examine two design choices that are underexplored: the loss used to compute gradients and how per-token attribution scores are aggregated into a document score. We find that the true causal evidence within a training document is highly localized, but standard methods dilute this evidence into noise by averaging attribution scores across the entire document. We find that replacing standard cross-entropy loss gradients with the negative log-odds gradients surfaces the critical tokens from other uninformative ones. We can then aggregate only the top token scores to improve ground truth recall by 35 to 72 percentage points across multiple baselines on our diagnostic benchmark. The negative log-odds correction also generalizes to FTRACE-TREX, improving counterfactual retraining performance for every method we test, by up to 4.9x.

In film editing, every frame matters. While macro, content-level edits shape the high-level narrative of a film, small local edits often control timing and direct visual attention. For example, editors can adjust the precise timing of clips to control the length of a meaningful silence, or cut to shots that highlight a specific character's reactions to important moments. These small, detailed edits are slow and tedious to make in standard editing tools because users have to map high-level narrative goals to low-level frame-by-frame adjustments. At the same time, these fine-scale adjustments, which we call micro-edits, can be nuanced and expressive, with different choices serving different creative goals. In this work, we start by conducting a formative study with professional editors to understand the structure of micro-edits and how they are used. We then derive a mathematical formalization for this structure to develop tools that leverage AI and optimization to make exploring micro-edits simpler and more efficient for users. Finally, we validate our tools in multiple user studies and a wide range of video examples.

The Neural Thickets hypothesis posits that pretrained language models lie in dense neighborhoods of task-specific experts in weight space, which random Gaussian perturbations can uncover. We reproduce the reported gains, but show that they do not require this geometric explanation. Perturbing a single decoder block already recovers much of the improvement, while keeping the model entirely frozen and adding a random bias to its output logits matches or exceeds full-weight perturbations on most reasoning tasks. This suggests an alternative explanation based on output diversity, validation-based selection, and voting. We formalize this view as Perturb-Select-Vote (PSV), a general procedure that generates perturbed predictors, selects high-performing members on a small validation set, and aggregates their predictions. This interpretation connects Neural Thickets to established inference-time methods, including sampling, self-consistency, and verifier-guided selection. A first-order analysis further shows that small weight perturbations act primarily through their induced logit shifts. We extend this analysis to vision, where logit PSV substantially improves performance under domain shift, including on Earth observation, while full-weight perturbations provide little additional benefit. Together, these results suggest that the gains attributed to nearby weight-space experts can instead be obtained through output diversity, selection, and voting. Logit PSV realizes these benefits with a fully frozen model, providing a simple, training-free alternative to full-weight perturbation ensembles across language and vision.

Pixel-space generative Transformers typically use one token per image patch, so capturing finer detail requires more tokens and makes global attention increasingly expensive. We observe that finer detail does not necessarily require more global communication: low-frequency structure benefits from global exchange, while high-frequency detail can be modeled locally. Based on this insight, we introduce SpecSlot, a pixel-space rectified-flow Transformer that decouples output detail from the length of the global attention sequence. SpecSlot expands each spatial anchor into multiple slots, each assigned a disjoint band of DCT coefficients; this separates anchor density from spectral subdivision. Only the lowest-frequency slot from each anchor participates in global attention, while all slots interact locally. This allows the model to increase spectral detail without enlarging the global sequence. A spatial head further adds a local correction to the spectral prediction, while the generative state, target, and loss remain in pixel space. On ImageNet-256, SpecSlot achieves a comparable final FID to a FLOPs-matched JiT control while converging faster. On ImageNet-512, it improves FID by 2.84 over the control with 3.9% fewer FLOPs, and comes within only 0.69 FID of JiT-B/16 at 38.2% lower cost. On text-to-image generation, SpecSlot surpasses JiT-B/8 on DPG-Bench by 1.06 at 40.6% lower forward cost.

Robot manipulation in everyday environments inherently involves substantial tool usage. Seemingly simple household tasks, such as cooking pan-fried tofu, require forceful cutting, delicate brushing, and precise placement using different tools under contact-rich interaction. Thus, the ability to autonomously swap and use task-appropriate tools is important for tool-centric manipulation. We introduce SwapBot, a modular, tool-centric end effector system featuring low-cost adapters that quickly convert common household tools into robot end effectors. Rather than pursuing a single universal gripper, SwapBot adopts a unified design principle that combines a fixed robot-side coupling interface with low-cost, tool-side adapters, enabling compatibility with diverse tools, including articulated ones, and multiple robot platforms. Each adapter integrates compact tactile sensors, supports fast autonomous swapping, and interfaces with an intuitive teleoperation system for efficient demonstration collection. User studies show that, during teleoperation, SwapBot reduces task completion time and operator workload. Using collected demonstrations on the robot, we train imitation learning policies to successfully perform tool-centric tasks like cooking tofu, stirring salad, flipping meat. Our ablation study shows that using tactile-augmented demonstrations improves policy performance across contact-rich skills, increasing success rates by 15 to 25 percentage points over RGB-only policies. These results highlight modular, tactile-equipped tools as a practical and extensible pathway toward robust and capable robot manipulation. Project website https://personalrobotics.github.io/SwapBot.

Distributionally robust optimization (DRO) hedges data-driven two-stage stochastic programs against error in the distribution estimated from a finite record, and standard ambiguity sets differ in how they treat realizations absent from that record. A divergence ball around the empirical distribution contains only reweightings of the record, whereas a transport ball also contains distributions supported off the record, including on realizations the data-generating process cannot produce, against which the decision is then hedged as well. We instead propose a Kullback-Leibler (KL) ball centered at a density model fitted to the record, the learned nominal, which lets the adversary up-weight costly unobserved realizations in proportion to the probability the model assigns them. Solving the resulting problem poses a second difficulty, because it can accommodate only a few scenarios, and an independent sample of that size is likely to omit the rare realizations on which the worst case concentrates. We therefore build the scenario set adaptively, alternating between solving the finite problem and drawing new realizations from the learned nominal, steered toward the worst case at the current decision, with importance weights that correct the draws back to the model. On a multi-item newsvendor problem and a capacity expansion problem on realistic grid data, centering the ball at the learned nominal rather than at the record lowers cost under distribution shift, at comparable in-distribution cost, when the record is scarce and the model class captures structure that lets it extrapolate beyond the record. The advantage fades as the record grows or as the shift departs from the structure the model has learned.

Luggage re-identification (ReID) is a critical task in video surveillance for forensics, security or border management (suspicious or abandoned luggage recognition), as well as luggage management (lost luggage recovery). It enables the association of observations of the same luggage across disjoint camera views in complex environments such as airports, train stations and border crossing. While existing approaches primarily focus on appearance-based cropped image matching under relatively clean conditions, real-world scenarios are characterized by frequent occlusions, cluttered background and crowded scenes which significantly degrade performance. Yet, strategies leveraging the complementarity of multiple shots and views from different video streams is overlooked. In this paper, we study strategies for an end-to-end pipeline that operates from multiple video streams to detect and re-identify luggage. We evaluate different variants on a benchmark dataset acquired in real-world environments to address the questions of selecting and merging different views and shots of the same luggage as well as training on different environment datasets for an overall ReID performance improvement in practical applications.

Extreme weather is forecast days ahead, yet its consequences for agriculture remain hard to predict: a forecast says where a storm will strike, not which crops will suffer or how much is at risk, and the records that could tell span many sources. We present WxCrop, the first agent that judges agricultural hazards from physically interpretable weather indices examined by causal analysis, and the first that combines weather and agricultural information to predict crop loss reliably for past and forecast dates. We demonstrate it in Florida on hurricanes, freezes and dry heat, with a loss model calibrated on insurance records for each, and the same procedure extends to other hazards and areas. Three checks show reliability: the hurricane index detects recorded cyclone days with an AUC of 0.98, recorded losses in held-out years are reproduced to within 2%, and independent production records confirm the county rankings (rank correlation 0.63). Users ask the agent for each crop group's loss as a fraction of expected season value, every number sourced, so that governments, growers and scientists respond to agricultural disasters quickly.

Recent progress in visual scene generation has extended beyond static 3D scenes toward the more challenging problem of generating dynamic 4D scenes from minimal input. However, sound, despite being intrinsically coupled with dynamic objects and events, remains largely unmodeled in prior work. We introduce Video-to-4DAVScene, a new task that turns a video into an explorable, time-varying 4D audio-visual scene, and present the SonoWorld4D framework for this task. Given a monocular video, our method generates a 360 static background, reconstructs explicit motion trajectories of dynamic, sound-emitting foreground objects, and synthesizes a physically-grounded sound field consistent across multiple sources through dual-space audio diffusion synchronization. The resulting 4D scene supports real-time novel-view visual and spatial audio rendering. Evaluation on real-world dynamic scenes and a user study demonstrates SonoWorld4D generates 4D audio-visual scenes with perceptually realistic, physically grounded spatial audio consistent with scene geometry and visual content.

World models can achieve strong control performance without faithfully capturing the dynamics of their environment, limiting their potential for mechanistic interpretability and scientific discovery. Here, we introduce world2vec to learn mechanistic world models without reward information directly from observations and actions. World2vec learns latent dynamics that faithfully reproduce the true system over long prediction horizons, with at least one order of magnitude fewer parameters than DreamerV3, TD-MPC2 and LeWM. This enables state of the art performance in controlling continuous environments from pixels and accurate system identification across a benchmark of 71 attractors, demonstrating how world models are amenable to scientific discovery.

The classic duck-rabbit illusion reveals a fundamental property of perception: when visual evidence is ambiguous, a visual system must resolve competing interpretations. But do humans and machine classifiers resolve the same ambiguity in the same way? To explore this, we use semantically ambiguous images as probes of human-machine perceptual alignment. Our psychophysically-informed framework interpolates between concepts in vision-language embedding space to generate continuous spectra of ambiguous images, enabling direct comparison of human and machine perceptual decision boundaries and sensitivities. We find systematic differences in these boundaries: machine classifiers are shifted relative to both human observers and the midpoint of the synthesis continuum. Guidance scale affects both bias and sensitivity in humans and the machine classifiers tested. Controlled semantic ambiguity in generated images, combined with a psychophysical framework, can thus enable quantitative evaluation of human-machine perceptual alignment.

An observer sees only part of the surrounding world, yet must keep track of what lies beyond the current view. Spatial perception therefore involves more than recognizing individual images: it calls for a persistent spatial state that is updated under self-motion and used to guide action. Do visual foundation models provide representations that support such a state? We study this question through four levels of spatial capability---spatial structure, object persistence, predictive spatial update, and actionability---using frozen-feature probes on ten representations under streaming egocentric observations of static indoor environments. The resulting capability profiles vary across tasks, network layers, and prefix lengths: strong readout of the visible scene does not ensure accurate object persistence or future-feature retrieval. Matched controls separate readout accuracy from reliance on visual history and motion. Native future prediction retains recoverable spatial structure, but a spatial reader trained on real features does not transfer to predicted features unchanged. Finally, goal-dependent candidate selection does not establish reliable continuous action utility: predicted actions respond to goal changes, yet supplying the correct goal does not reliably improve progress. Across the four levels, object persistence and actionability build on camera-pose readout, whereas feature prediction builds on visible-scene readout. Together, these findings motivate evaluating spatial representations not only by what can be decoded from them, but by whether that information is maintained beyond observation, updated predictively, and used for goal-directed action.

As robots and intelligent systems increasingly interact with humans, they must anticipate long sequences of future actions from only brief observations. While purely neural models typically require extensive observations to predict future actions and remain difficult to interpret and adjust, symbolic priors offer a structured alternative. To this end, we introduce SymAnt, a novel neuro-symbolic approach that unifies the expressiveness of neural networks with the structured reasoning of scene and knowledge graphs. Given a short video, SymAnt performs a joint-search over both graphs to construct a symbolic context, enabling unified reasoning over spatial relationships, object attributes, affordances, and associated actions. This symbolic context serves as a prior for an iterative refinement process that progressively generates and improves the sequence of future actions. By grounding predictions in symbolic context, SymAnt anticipates long-horizon actions from short observation windows in a parameter-efficient manner, while enabling users to directly edit the underlying knowledge to adjust predictions without further training. Through our experiments, we demonstrate the effectiveness of SymAnt, particularly when only short observation contexts are available, across five diverse datasets: Breakfast, 50 Salads, EPIC-KITCHENS, EGTEA Gaze+, and Assembly101.

Safety alignment has enabled large language models to reliably refuse harmful requests directly, yet the same request distributed across a multi-turn conversation still succeeds far too often. This brittleness has a mechanistic root: Alignment is shallow, with safety behavior concentrated in a model's first few output tokens, and existing defenses inherit the weakness, supervising the surface form of the response without ever giving the model an explicit safety decision to learn. We take the opposite view of shallowness: if a short span determines the trajectory of an entire response, that span is the highest-leverage place to intervene, and the defender can occupy it first. We propose Assess $\to$ Decide $\to$ Respond, a framework in which the defender assesses the user's intent, commits to an explicit, parseable safety decision, and only then generates a reply conditioned on it, with the commitment trained by reinforcement learning at the payload turn. Because the decision is a token span, the reward can be a rule: a self-verified reward checks the committed label against the conversation's known nature, requiring no judge and no annotation, while the same format also supports an LLM-as-judge reward over a graded decision space with safe redirection. Across two model families and fourteen benchmarks spanning safety, utility, and helpfulness, our defenders achieve the strongest multi-turn safety in a controlled comparison, maintaining their lead in safety while preserving utility at a modest cost to helpfulness. A per-token KL analysis and interventional probes confirm why this works: training concentrates its effect on the brief deliberation that sets a response's trajectory, and the decision committed there is a causal control point. The very shallowness attackers exploit becomes the mechanism of defense.

Adapting Large Language Models (LLMs) to specific tasks usually involves fine-tuning: Low-Rank Adaptation (LoRA) edits the weights, while soft prompting prepends trainable virtual tokens. Both change how an already-optimized model is served, so serving many task-specific adapters at high throughput requires adapter-aware infrastructure. We propose Art-Based Reinforcement Training (ART), which optimizes a single auxiliary input image to specialize a frozen Multimodal Large Language Model (MLLM) for a specific task. At inference time, ART simply includes this image in each standard multimodal prompt, requiring no changes to the serving engine. We compare ART with LoRA and soft prompting across five reasoning benchmarks for four small MLLMs from three architectures. While different methods perform best for different tasks, ART ranks first or second in 10 of 20 benchmark-MLLM pairs, with its best results on the two math word problem benchmarks (GSM8k and SVAMP). It is significantly ahead of the next best adapter in 2 cases, within the noise of the next best adapter in 6, and significantly above the text-only baseline in 18. ART therefore reaches accuracy close to LoRA and soft prompts while serving faster than LoRA. In the vLLM serving stack, ART reduces throughput by 6% to 14% relative to text-only requests, compared to LoRA's 13% to 93% slowdown. Beyond efficient serving, ART artifacts are plain images, making them easy to store, distribute, and share. This approach keep task-specific adaptation outside the model weights, so the same frozen model can support different tasks without separate model variants.

Repository-level software engineering (SWE) comprises heterogeneous task categories, whose progress under pooled agentic reinforcement learning can be uneven: gains in some categories coincide with regressions in others, while aggregate resolution obscures these changes. Motivated by this category see-saw, we develop a category-aware expert-training and policy-integration framework. Executable task construction and SWE Labeler, an evidence-grounded multi-axis labeling system, organize the training pools. Initial category-specific RL improves average training success while leaving uneven instance-level progress, motivating explicit consolidation of successful behavior and policy-adaptive task selection. Same-origin category experts alternate long-horizon Agentic-miniRL with Refresh-Repair-Expand (RRE): the updated policy refreshes instance mastery, reuses its own verified successful trajectories for Repair SFT, and reselects tasks for further RL. Label-routed multi-teacher on-policy distillation (MOPD) consolidates the experts into one deployable student, with ReLU-gated reward extrapolation keeping only each teacher's improving direction over the reference. Expert training and policy integration require no external model to provide solution trajectories or action targets. On a fixed, audit-filtered Pro-618 subset, we connect Pooled RL and Balanced RL, expert development, and single-model integration through overall and per-category resolution and expert-gain recovery. In final-model evaluations, MOPD achieves mean resolution of 86.19% on SWE-bench Pro V2 and 71.78% on SWE-bench Multilingual, exceeding the base model by 4.57 and 3.89 percentage points, respectively. The same three 27B experts also improve a Qwen3.6-35B-A3B student across the evaluated task sets.

Differential privacy is the leading mathematical framework for privacy protection, providing a probabilistic guarantee that safeguards individuals' private information when publishing statistics from a dataset. This guarantee is achieved by applying a randomized algorithm to the original data, which introduces unique challenges in data analysis by distorting inherent patterns. In particular, causal inference using observational data while preserving privacy is challenging because it requires a good covariate balance between treatment groups, but checking the covariate balance is challenging when privacy is a primary concern because releasing any statistic (e.g., t-statistic) for the balance check compromises privacy. Additionally, the performance of the privatized estimator critically depends on the choice of differential privacy mechanisms, and it remains unexplored how privacy-protecting causal inference could be conducted with observational data. In this article, we present a differentially private two-stage covariate balancing weighting estimator to infer causal effects from observational data. Our algorithm produces both point and interval estimators with statistical guarantees, such as covariate balance, rate optimality, and oracle budget allocation, under a given privacy budget.

Ask a language model whether sales tend to dip after a product recall, or whether one side effect tends to follow another after a vaccine, and it answers in a fluent sentence that reads like a finding, with no test behind it. Assertions about how events unfold across many cases fill the Web, and each is a temporal hypothesis that only a significance test over the right records can settle. The verifier must specify the null, select the test and decide which records bear on the hypothesis, and any of these choices can go wrong while the conclusion still comes out right. Yet benchmarks for language-model data analysis grade only the conclusion, with the relevant data handed over. We study the problem in health, where a wrong verdict is costly, and present STRATA, a benchmark that grades every decision on the way to a verdict: 8,735 hypotheses over the clinical timelines of vaccine adverse-event reports, with gold plans, evidence sets, p-values and verdicts derived under a released codebook and reproduced by a deterministic oracle. We also propose ART (articulate, retrieve, test), a staged paradigm that exposes each decision as an inspectable output. Extensive experiments with five model configurations show that systems fail mostly at choosing which records count, a verdict-evidence gap that accuracy alone conceals. ART closes most of this gap and brings a 31B open model to within 0.01 of the frontier models under the same paradigm; the frontier models succeed once given the codebook.