Self-distillation (SD) offers a promising path for adapting large language models (LLMs) without relying on stronger external teachers. However, SD in autoregressive LLMs remains challenging because self-generated trajectories are free-form, correctness is task-dependent, and plausible rationales can still provide unstable or unreliable supervision. Existing methods mainly examine isolated design choices, leaving their effectiveness, roles, and interactions unclear. In this paper, we propose UniSD, a unified framework to systematically study self-distillation. UniSD integrates complementary mechanisms that address supervision reliability, representation alignment, and training stability, including multi-teacher agreement, EMA teacher stabilization, token-level contrastive learning, feature matching, and divergence clipping. Across six benchmarks and six models from three model families, UniSD reveals when self-distillation improves over static imitation, which components drive the gains, and how these components interact across tasks. Guided by these insights, we construct UniSDfull, an integrated pipeline that combines complementary components and achieves the strongest overall performance, improving over the base model by +5.4 points and the strongest baseline by +2.8 points. Extensive evaluation highlights self-distillation as a practical and steerable approach for efficient LLM adaptation without stronger external teachers.
Hao Li, Zhengkun Zhang, Gangqiang Hu, Zhen Zhang, Yude Gao, Dai Dai, Jing Liu
Reinforcement Learning with Verifiable Rewards (RLVR) is expanding from tasks with well-defined correctness signals, such as mathematics and code, toward multifaceted quality requirements specified by multi-dimensional rubrics. Since policy optimization consumes one scalar per rollout, rubric-based pipelines must map multiple criterion scores into a scalar reward. This aggregation is often treated as score scaling, but it implicitly determines how quality dimensions trade off during training. The prevailing practice, normalizing each criterion and taking a linear combination, assumes that cardinal score differences are comparable across criteria and that gains on one criterion compensate for failures on another; both assumptions are unreliable when criteria are semantically heterogeneous. We propose Reinforcement Learning with Verifiable Rubric-based Ranking (RLVR$^2$), a verifiable ranking paradigm for rubric-based RLVR. For each criterion, RLVR$^2$ converts rubric scores into criterion-specific within-group ordinal outcomes, recovers a latent utility from the resulting comparison matrix, and merges these utilities into one training signal. By retaining only within-group ordering and discarding raw score magnitudes, RLVR$^2$ avoids calibrating heterogeneous rubric scales. It further supports objective-preserving attribute adjustment: auxiliary attributes that correlate with observed rankings but are not training objectives can enter the estimation without expanding the rubric or rewarding them directly. Across three model scales and 16 benchmarks, RLVR$^2$ consistently outperforms representative rubric-based baselines, achieving the best overall performance on most benchmarks at every scale. Analysis shows it controls systematic effects tied to reasoning efficiency and response formatting while preserving the quality objective.
Articulated 3D assets are essential for robotic simulation and training, yet their generation from casual images poses a key challenge: recovering part geometry and kinematics that remain mutually consistent. We present ArtCrafter, a novel feed-forward generative model for controllable generation of articulated 3D objects from a pair of images, one depicting the object in the rest state and the other in the open state. At the core of ArtCrafter is a DiT-based generative model that jointly produces part-level meshes for both articulation states in a single forward pass. Inspired by Slot Attention, its denoising process operates on a $1+2N$ slot layout comprising one shared base slot and $N$ paired rest/open slots, encouraging each pair to bind to the same kinematic part. To ensure geometrically consistent and disjoint part generation, we introduce Triplet Attention to enforce cross-state geometric consistency between paired rest/open slots anchored through the shared base slot, and a repulsion loss $\mathcal{L}_{\mathrm{rep}}$ on decoded part point clouds to prevent inter-part penetration throughout denoising. Given the resulting paired meshes, URDF joint parameters---type, axis, origin, and range---are derived analytically from the rigid transformation between each rest/open pair, providing an interpretable alternative to learned joint-parameter regression. Extensive experiments on both in-distribution and out-of-distribution datasets demonstrate that ArtCrafter outperforms existing method under the same dual-state setting in geometry quality and articulation precision, while remaining competitive with methods operating under different input settings and paradigms.
The proliferation of computation-intensive applications, such as generative artificial intelligence and low-altitude economy services, demands mobile edge computing (MEC) systems that are both energy-efficient and adaptable to highly dynamic workloads. However, conventional MEC architectures rely on fixed-position antennas with limited spatial flexibility and rigid single-slot task execution models, which fail to accommodate bursty traffic and time-varying channels. To address these challenges, this paper proposes a novel MEC framework empowered by movable antennas (MAs) and task buffering. Unlike traditional approaches, our system enables continuous antenna repositioning to maximize array gains and decouples task generation from immediate execution via buffer queues, allowing for flexible multi-slot scheduling. We formulate a stochastic optimization problem to minimize long-term energy consumption by jointly optimizing receive beamforming, MA positions, and computing resources. Leveraging Lyapunov optimization, we transform the long-term problem into tractable per-slot subproblems and develop an efficient alternating optimization algorithm. Simulation results demonstrate that the proposed scheme significantly outperforms the benchmark schemes in terms of energy consumption, while maintaining stable task queues under unpredictable traffic conditions.
Online conformal prediction (OCP) seeks prediction intervals that achieve long-run $1-α$ coverage for arbitrary (possibly adversarial) data streams, while remaining as informative as possible. Existing OCP methods often require manual learning-rate tuning to work well, and may also require algorithm-specific analyses. Here, we develop a general regret-to-coverage theory for interval-valued OCP based on the $(1-α)$-pinball loss. Our first contribution is to identify \emph{linearized regret} as a key notion, showing that controlling it implies coverage bounds for any online algorithm. This relies on a black-box reduction that depends only on the Fenchel conjugate of an upper bound on the linearized regret. Building on this theory, we propose UP-OCP, a parameter-free method for OCP, via a reduction to a two-asset portfolio selection problem, leveraging universal portfolio algorithms. We show strong finite-time bounds on the miscoverage of UP-OCP, even for polynomially growing predictions. Extensive experiments support that UP-OCP delivers consistently better size/coverage trade-offs than prior online conformal baselines.
Expert pruning compresses Mixture-of-Experts (MoE) language models by removing the experts that a calibration set marks as unimportant. Studies of high-resource languages largely find that English calibration transfers well to other languages. We test this for Lithuanian: we pruned four MoE models with REAP, removing 25% or 50% of the experts with English, Lithuanian or mixed calibration, and evaluated both languages on reading comprehension, translation and language modelling. English calibration does not transfer to Lithuanian: at 25% pruning, English comprehension fell by a median of 1.7 points but Lithuanian by 21.4. The transfer is asymmetric. Lithuanian calibration removed most of the Lithuanian loss and never significantly hurt English comprehension, which was on average 2.5 points better than under English calibration; only English language modeling got slightly worse. Routing suggests why: in the last layers, Lithuanian sends 55-72% of its expert selections to just eight experts, against 19-39% for English, and English calibration removes some of these experts. Compressed models should be calibrated, and evaluated, in every language they must serve.
Haichuan Tan, Bowen Gao, Tao Huang, Jiaxin Tan, Xin Hong, Baoguo Yao, Yifei Wang, Zhichao Deng 等 25 位作者
Drug discovery requires continuous scientific decision-making across a vast research space. We introduce PharmAgents2, an autonomous system for end-to-end drug discovery that combines knowledge-enhanced decision-making with organized large-scale exploration. Paper2Skill extracts transferable scientific decision knowledge from public sources and, together with expert-developed skills, supports research decisions. Self-Scaling Discovery expands exploration through parallel agents inheriting research context and decision-making capabilities across targets, scientific hypotheses, and discovery strategies. We introduce PharmAgents-Bench v0.1, comprising 121 questions across nine decision scenarios and three end-to-end tasks assessed with expert-defined process rubrics. PharmAgents2 outperforms evaluated baselines in most decision categories and achieves the highest process scores on all three end-to-end tasks. In three real-world studies, autonomous investigations generated findings subsequently supported by human-executed experiments. For LpxH, PharmAgents2 hypothesized a previously unreported allosteric site and screened compounds against it, identifying a novel inhibitor scaffold with experimentally confirmed inhibitory activity and binding. For JAK1, it prioritized a covalent inhibition strategy and nominated compounds with experimentally demonstrated activity. Starting from a known bioactive molecule, Self-Scaling Discovery explored candidate targets across the proteome and identified a new binding target, with the interaction subsequently confirmed experimentally. PharmAgents2 represents a prototype of a virtual pharmaceutical company, marking a step toward autonomous end-to-end drug discovery.
Current autonomous logical systems rely on a fragile assumption: that language models should dynamically write code, guess API payloads, and memorize operational constraints through prompt engineering. In practice, this burns massive context windows, inflates API bills, and produces brittle agents that fail unpredictably in production. In this work, we present Skillware, an open architectural framework and capability registry that formally separates what an agent wants to do from how that task is executed. Instead of prompting models with loose text instructions or markdown files baptized as 'skills', Skillware equips them with modular, versioned software capability bundles. Much like skill hotkeys in video games that execute deterministically under fixed world rules, each bundle combines a typed JSON input contract, deterministic Python code, task directives, bundled datasets, and automated unit tests. In-process adapters translate these capabilities into native tool declarations for any model, from frontier cloud APIs to local open-weight runtimes with zero network latency. Evaluating 23 frontier and open-weight models across 368 benchmark conditions, we show that small local models (3B to 8B parameters) equipped with Skillware consistently outperform unequipped frontier models on complex multi-step workflows. They cut token usage by up to 84.8%, slash inference costs by up to 97.3%, and eliminate parameter errors entirely. True agent reliability does not come from larger models or longer prompts, but from equipping logical systems with verified, deterministic software. Finally, we articulate the trajectory of capability bundles across an invariant continuum of information processing systems: from digital logical systems today, to IoT, drones, and embodied robotics tomorrow, establishing a foundation for auditable, real-time capability execution.
Modern 6-DoF grasp generators are often evaluated using grasp scores, dataset-level metrics, or single-robot demonstrations, yet proposal quality alone does not determine whether a grasp is executable by a particular robot. We present GraspBench, a controlled MuJoCo protocol for evaluating complete grasp-generation pipelines after robot-specific inverse-kinematics filtering, motion planning, and fixed approach-close-lift execution. GraspBench evaluates six configured pipelines across four robot embodiments, five diagnostic objects, and a structured tabletop pose grid, yielding 540 scenarios per method and 3,240 method-scenario evaluations in total. Missing outputs are conservatively counted as failures, and success follows the binary lift criterion produced by each method's execution runner. GraspGen achieves the highest overall success rate (73.1%), followed by EconomicGrasp (63.5%), while robot- and object-conditioned analyses reveal substantial variation across embodiments and object geometries. GPG produces the largest native candidate sets but achieves 44.3% success, indicating that proposal volume alone does not guarantee robot-executable grasps. Cluster-aware paired intervals support the principal method differences while not estimating repeated-execution uncertainty. The results characterize complete configured pipelines on fixed scenarios rather than intrinsic model quality or stochastic repeatability, showing that grasping performance depends jointly on the generator, robot embodiment, scene, and execution protocol.
Yonathan Wolanowsky, Guy Bar-Shalom, Fabrizio Frasca, Haggai Maron
Weight-space learning treats the weights of neural networks as data, processed by other models called metanetworks. A key challenge is that different weights can realize the same function, related by a group of symmetries the metanetwork should respect. While symmetry-aware metanetworks for simple architectures such as MLPs are well developed, a comprehensive treatment of transformers, and in particular transformer-based language models, is still missing. Importantly, their weight space has a richer symmetry group: beyond neuron permutations, attention introduces continuous general-linear symmetries of the attention matrices and permutations of entire heads. Existing methods for transformer weights either capture only part of these symmetries or are restricted to a single, fixed architecture, and none accounts for language-model-specific components such as the token embeddings. Building on graph metanetworks, we propose a graph-based approach that is both symmetry-aware and applicable across depths and widths. We formalize a weight-space symmetry group of complete decoder-only language models and provide a graph construction that is invariant to it. We further introduce two new transformer-population datasets with novel tasks, such as training-corpus prediction and fine-tuning detection, on which our approach notably outperforms existing baselines.
Third-party application programming interface (API) relays offer access to large language models (LLMs), often at lower prices. Users depend on these services for model quality, instruction compliance, and accurate billing. However, existing studies leave unclear whether relays meet these expectations together. We test 20 relays offering GPT-6 Astra and Claude Opus 5. First, we compare answers to knowledge questions with batch-matched official responses. In our first-round sample, the analysis flags 7/10 low-price and 2/10 same- or higher-price relays for behavioral differences. On retesting 35 pairs, seven gain or lose a flag under the original rules, and six under matched correction families. Because answer comparisons cannot reveal every delivery problem, we examine returned content and tool permissions. No injection is confirmed in 3,332 screened responses to a fixed, non-adversarial code and setup workload, yet four relays return tool calls despite instructions disabling them in the first round. Finally, to connect reported usage to actual costs, we compare input-token counts with official references and check balances. For identical requests, five relays report 4,200 to 60,000 more input tokens than the official API. Where balances are available (six services), charges follow reported usage at listed prices, including extra reported input tokens at one service. Our study shows that advertised models and prices are insufficient to judge relay reliability. It identifies unexplained billed token quantities and tool-permission violations as risks and shows why service evaluations need repeated measurements.
Central to human-aligned AI is understanding the benefits of human-elicited labels over synthetic alternatives. While human soft-labels improve calibration by capturing uncertainty, prior studies conflate these benefits with the implicit correction of mislabeled data (mode shifts), obscuring true effects of soft-labels. We present a controlled audit of soft-label learning across MNIST and a synthetic variant, re-annotating subsets to extract human uncertainty. By decoupling soft-label supervision from underlying label mode shifts, we show that while human soft-labels do provide accuracy gains, their larger value lies in acting as a regularizer that improves model calibration on difficult samples and promotes stable convergence across training runs. Dataset cartography reveals models trained on human soft-labels mirror human uncertainty, whereas those trained on synthetic labels fail to align with humans. Broadly, this work provides a diagnostic testbed for human-AI uncertainty alignment.
Ziqiao Weng, Yuanzhi Liang, Haibing Huang, Weidong Cai, Xuelong Li
Post-training aims to enhance desired attributes in visual generative models while preserving their general generation capabilities. Reward-guided post-training faces two challenges: real training data may mismatch the generator's output distribution, and attribute supervision may be confounded by unrelated content. Unbounded reward maximization can also cause overoptimization and degrade generation quality. We propose a framework combining Generator-Aligned Reward Modeling (GARM) and Target-Aware Bounded Optimization (TABO). GARM mitigates distribution mismatch through generator-support matching and reduces content confounding using content-matched real contrasts, yielding generator-relevant, attribute-focused reward supervision. TABO maps the learned reward to a smooth, upper-bounded function calibrated from real target-positive examples. This transformation is independent of the optimization algorithm and supports both direct reward-gradient and policy-based optimization. We further introduce Target-Aware GRPO (TA-GRPO), an adaptation of Flow-GRPO that combines fixed stage-wise advantage scaling, a generator-to-target curriculum, and timestep-balanced credit assignment for flow-based video optimization. Experiments show that GARM improves target-attribute expression while maintaining competitive generation quality. TABO better balances attribute enhancement and visual quality than unconstrained reward maximization, while TA-GRPO improves target enhancement over vanilla Flow-GRPO and reduces unintended attribute leakage on neutral prompts. These results highlight the importance of generator-aligned reward supervision and controlled, data-calibrated optimization for targeted visual generation.
In this paper, we introduce Universal Generator-Agnostic Deepfake Detector (UniGADD), a novel method that employs supervised learning to achieve high accuracy on known deepfake generators, while maintaining robust performance on previously unseen ones. The proposed approach follows a two-stage optimisation process. In the first stage, a contrastive loss encourages the model to learn discriminative feature embeddings from real and known fake images, resulting in strong performance within the training domain. In the second stage, the embedding space is refined by promoting inter-cluster separation and intra-cluster compactness, applied exclusively to real samples. This refinement enhances generalisability, enabling the method to exhibit improved robustness against unseen deepfake generation techniques. UniGADD achieves accuracy on par with state-of-the-art methods for known generators, while significantly outperforming them on unseen cases, demonstrating its scalability and practical applicability for adversarial content detection.
Ryan Liu, Chuwei Wang, Ricardo Baptista, Andrew Stuart, Anima Anandkumar
Neural operators offer fast surrogates for solving PDEs, but training them typically requires expensive input-solution pairs. Two solver-free alternatives include physics-informed learning and training on manufactured data. Physics-informed learning approximates solutions by minimizing PDE residuals. While the residual norm must be strong enough to control solution error, stronger norms increase mini-batch variance and hinder optimization. Manufactured data constructs input-solution pairs for learning by applying the PDE operator to artificial solutions. However, the resulting training distribution can be mutually singular with the target distribution, yielding poor accuracy on the target distribution. We introduce the Wasserstein Data Synthesis Pipeline (WaSP), which trains on manufactured input-solution pairs while gradually mitigating the distribution shift. WaSP synthesizes new pairs by applying the learned operator to a path of intermediate distributions converging to the target input distribution. Training proceeds in stages along this path. At each stage, the previous surrogate generates solutions, and the exact PDE operator derives corresponding inputs, yielding training pairs for the next surrogate. Theoretically, we prove a stepwise error-propagation bound for the distribution shift introduced during synthesis. Empirically, WaSP reduces solution error by $3.9\times$ to $9.5\times$ compared to existing solver-free methods on PDEs such as Allen-Cahn, Kuramoto-Sivashinsky, and Ginzburg-Landau, and better reproduces the two chaotic systems' statistical observables.
Existing context models for learned image com- pression rely primarily on internal priors from the current image, overlooking common patterns and typical textures shared across natural images. To address this limitation, we propose an Internal-External Prior Fusion-Alignment (IE-PFA) context model built upon quadtree spatial-channel joint modeling, where a learnable dictionary shared by the encoder and decoder serves as an external prior. Specifically, an RWKV-based Spatial- Channel Prior Fusion (RSCPF) module efficiently fuses inter- nal and external priors along spatial and channel dimensions with linear complexity, while a training-only Self-Distilled Prior Alignment (SDPA) mechanism aligns them without inference overhead. Moreover, our model supports variable-rate coding over 64 quality levels within a single network and is accelerated by custom CUDA operators. Experiments on seven benchmarks demonstrate state-of-the-art rate-distortion performance, achiev- ing an average BD-Rate saving of 21.65% over VTM-22.0-Intra, with only 29.9M parameters and the fastest coding speed among advanced neural codecs (61 ms encoding and 52 ms decoding). Code is available at https://github.com/stcnjtech/DCVC-X-Intra.
This paper proposes a continuous-time, projection-free neurodynamic system based on gradient tracking for distributed nonnegative sparse optimization. Unlike traditional methods, our approach avoids projecting states onto the feasible set. Instead, it employs gradient tracking to estimate average gradients under heterogeneous costs, coupling primal and dual dynamics such that feasibility and sparsity emerge naturally from the equilibrium conditions. The system ensures unique solutions through a locally Lipschitz vector field and guarantees that every equilibrium corresponds to an exact KKT point. We establish rigorous convergence guarantees for convex objectives through three distinct analytical layers. These include a centralized kernel rate, a computable distributed rate derived from Jacobian analysis without restrictive stability assumptions, and a global exponential stability proof based on Lur'e system theory. Our analysis quantitatively separates these layers, revealing that the convergence rate is primarily limited by graph sparsity and remains robust to variations in momentum time scales. Numerical experiments on strongly convex quadratic problems validate the theoretical rates, demonstrate effective active set recovery, and show that the system maintains stable tail convergence rates even under strong local heterogeneity.
Sebastian Schmidt, Juha Harviainen, Corentin Moumard, Aleksandr Politov, Francisco Sena, Alexandru I. Tomescu
In recent years, in attempts to capture variation structures in genomic and pangenomic graphs, the zoo of bubbles has been filled with more and more different variants. The first more complex variant was the superbubble which was later (informally) generalised to bidirected graphs as the ultrabubble. Ultrabubbles are more applicable in bioinformatics research, because genome and pangenome graphs usually are bidirected. In this paper, we study the relationship between superbubbles and ultrabubbles, and show that ultrabubbles in a bidirected graph are equivalent to the weak superbubbles in the corresponding double covering graph. Using this, we introduce a new linear-time algorithm to enumerate all ultrabubbles that consists of a very simple reduction to employ an existing linear-time algorithm to enumerate all weak superbubbles. Treating the superbubble algorithm as a black box, our algorithm is vastly simpler than the SPQR-tree-based linear-time ultrabubble enumeration algorithm of Sena et al. [2026]. Further, since our algorithm works on all bidirected graphs, it is strictly more powerful than the simple graph-orientation-based algorithm by Harviainen et al. [2026], which only works on a restricted class of bidirected graphs. In order to prove our equivalence, we additionally introduce simplified definitions of both ultrabubbles and weak superbubbles, through which the equivalence becomes more intuitive. Finally, our reduction motivates that some algorithmic problems in bidirected graphs can be reduced to well-known algorithmic problems in directed graphs.
Rongting Zhang, Martin Bertran, Sandeep Avula, Chia-Jung Lee, Vanessa Murdock
Indirect prompt-injection (IPI) benchmarks predominantly test planted instructions that redirect a tool-using agent toward an attacker-chosen task. Yet attacker-controlled content can instead tamper with the outcome of the task already in progress. A buyer's dispute evidence can cite a false refund policy to inflate a refund without changing the task. We define IPI by authority rather than instruction-like wording: planted content constitutes an attack when it attempts to influence a decision that its source is not authorized to determine. We introduce a benchmark that makes this boundary explicit: in each of 90 validated cases across three tool-use environments, the user request specifies the source that resolves the targeted decision. Every attack tampers with a decision required by the benign task rather than replacing that task. Each case has attacks with the same harmful outcome that target up to six parts of the agent's decision making: its objective, decision rules, available actions, expected consequences, reading of observations, or beliefs about the current state. Without a defense, attacks on three of the five other targets (decision rules, available actions, and state) have higher mean success than objective attacks across five LLMs. A detection-and-removal defense reduces objective-targeting attacks to 2.4% success but leaves observation- and state-targeting attacks at 9.5% and 12.1%; this gap persists under detection-optimized rewrites. A tool-policy defense still allows 15.4% attack success, close to 17.0% without a defense, because these attacks generally retain expected tools and manipulate values obtained during execution. Evaluations whose attacks redirect the task or target only the agent's objective can therefore overestimate the coverage of these defenses. Code and data are available at https://github.com/amazon-science/task-tampering-prompt-injection.
Reinforcement learning from verifiable rewards (RLVR) is usually evaluated as an alternative to supervised fine-tuning (SFT), or as a stage that follows it, with the implicit assumption that the RL stage contributes gains of its own. We measure that contribution directly, in a controlled setting where both stages can be run to completion and compared exactly. We built gpu-train-debug, a deterministic benchmark of 1,877 synthetic GPU and LLM training failure diagnosis tasks across 30 archetypes, graded by a programmatic verifier with partial credit, and calibrated so that a mid-tier model scores in a 0.60 to 0.75 band. A pilot produced a surprise: 1,200 gold demonstrations nearly solved the benchmark (SFT 0.785 on the hard split), while 15 steps of GRPO from the base model did not move it (0.314 vs 0.311 base). We therefore amended the study, in writing and before the full run, to ask the sharper question: what does verifier-based RL add beyond demonstrations? At the pre-registered scale (60 GRPO steps, learning rate 1e-6), the answer is: no detectable improvement in-distribution (+0.22 points, 95% CI [-0.06, +0.51]), and a modest but measurable improvement under distribution shift (+0.33 points on a held-out OOD split, CI [+0.02, +0.68]). SFT accounted for most of the available improvement, while verifier-based RL produced only modest additional gains under distribution shift. The residual errors SFT leaves are mostly systematic, identical across samples, which likely reduces the within-group learning signal available to GRPO. A full 60-step cold-start GRPO arm confirms the pilot at the full horizon. The whole study cost $8.69 of compute and is fully reproducible from archived generations, frozen configs, and tagged code.
Large language models (LLMs) increasingly assist crowdsourced fact-checking, yet evaluations focus on individual annotations. A correction's reach and reception depend on visibility, audience responses, and platform policies. We introduce correction context understanding: assessing corrective content, interpreting audience responses, and reasoning about platform dynamics within a shared environment. We build NoteHabitat from 36 months of Weibo Annotations data, linking 102,994 posts (42,853 annotations), 431,407 comments, and five policy events. Evaluating diverse model families across seven tasks spanning content, interaction, and management reveals three gaps. Annotation quality correlates only weakly with engagement ($ρ=0.068$), while four zero-shot LLMs systematically underrate quality relative to human judgments, especially evidence support and objectivity. Models also miss toxic comments supporting corrections (36.4% average recall across ten models). At the platform level, LLMs partially predict policy-associated outcome directions, but LLM-guided simulations poorly reproduce daily activity. Cross-group analyses expose blind spots tied to contribution types and reader stances that aggregate scores conceal. Reliable fact-checking AI must understand the note, the conversation around it, and the platform in which both unfold.
Yuqiu Jiang, Xiaozhen Qiao, Yifan Chen, Ye Zheng, Zhe Sun, Xuelong Li
Long-tailed human-object interaction (HOI) detection remains challenging because rare interactions are sparsely represented in training data. Although test observations provide additional visual evidence, existing cache-based methods typically assign uniform memory capacity across classes, limiting their ability to capture informative features of underrepresented interactions. To address this issue, we propose Adaptive Diversity Cache (ADC), a training-free, plug-and-play framework that adaptively constructs test-time caches through frequency-aware capacity allocation and confidence-diversity joint sample selection. Specifically, Frequency-Aware Cache Adaptation assigns larger capacities to underrepresented interactions using frequency-dependent priorities and a weighted logarithmic allocation objective, and fills remaining slots with selected augmented views. Confidence-Diversity Joint Cache Selection further combines prediction confidence with local feature diversity to retain informative cache entries. Cache retrieval then refines the predictions of the frozen detector. Experiments on HICO-DET and V-COCO across three base detectors demonstrate consistent improvements. On HICO-DET, ADC improves mean average precision (mAP) for rare interactions by 4.02--8.57 percentage points, while also improving performance on non-rare interactions. Online adaptation further achieves gains of 2.87--7.36 percentage points on rare interactions. These results demonstrate that adaptive test-time memory effectively complements existing HOI detectors and improves long-tailed detection performance.
Dong-Hee Kim, Seonwoo Choi, Changbeen Kim, Jungmyung Wi, Juyeon Ko, Youngju Choi, Il Hyeon Mun, Hyunwoo J. Kim 等 9 位作者
Understanding long-form video remains a fundamental challenge for multimodal large language models (MLLMs). Sparse frame sampling fails to capture fine-grained visual details, while dense sampling quickly exceeds context length limits. Retrieval-augmented generation (RAG) offers a promising middle ground by selectively retrieving relevant video segments for grounded generation, yet its effectiveness critically depends on the quality of the video segments used as retrieval units. In this paper, we investigate RAG for movie understanding, which demands story-level reasoning over characters, events, and narrative arcs spanning hours of content. Scene segmentation, a long-studied problem that partitions movies into semantically coherent units, is a natural candidate for defining such retrieval units. We reexamine whether existing methods actually serve this role through comprehensive evaluation on downstream movie understanding tasks, and find that they do not provide a consistent or dominant advantage over naive uniform temporal chunking. Our audit of the most standard scene segmentation benchmarks reveals why: current annotations prioritize visually salient transitions over narrative event structure. Motivated by this mismatch, we introduce NarraScene, a narrative-centric scene segmentation dataset annotated with a three-level cognitive taxonomy spanning physical, character, and narrative change, where every valid boundary requires a narrative-level shift. When used as retrieval units, these narrative-grounded segments outperform uniform chunking on downstream movie understanding tasks, suggesting that the central challenge for scene segmentation in movie RAG is not detecting boundaries, but identifying the narrative event units that matter for movie understanding. The dataset is available at https://huggingface.co/datasets/queez0405/NarraScene.
Large language models increasingly assist developers by proposing small code edits, yet reliably integrating a partial update snippet into an existing file while preserving unrelated content remains poorly supported by current code-generation workflows. We formalize the Code Integration (Apply) task as follows: given an original file and an update snippet, the system must produce the complete merged file while preserving all content outside the requested edit. We construct a high-quality multi-language dataset from real commits via a structured pipeline that converts code changes into update snippets paired with full-file ground truth and enforces consistency through verification and filtering. We then train a specialized 4B model, aiXapply, using supervised fine-tuning and reinforcement learning (GRPO). The supervised variant achieves 94.4% accuracy under our language-aware comparison, surpassing a strong large-model (Deepseek-V3.2) baseline (91.6%). Speculative decoding enables aiXapply-RL to reach $\sim$2700 tokens/s on a single A100 in a decoding subset; on the full benchmark, it achieves roughly 75$\times$ lower latency than the API-served Deepseek-V3.2 baseline. We further conduct an industrial deployment of aiXapply-RL in the aiXcoder IDE plugin as a dedicated Apply service, reducing mean Apply latency from 50 seconds to 1.89 seconds by offloading full-file generation from the upstream large model. Under time-based capacity planning, replacing one of five upstream serving instances would save 20% of upstream machine cost before the additional shared A100 service cost. Overall, our results demonstrate that treating code integration as a standalone task enables small models to achieve high accuracy, practical efficiency, and concrete industrial benefits in real-world IDE workflows where latency, cost and privacy are critical.
Sai Yashwant, Manoj Kumar Gondimalla, Arjun Sharma, Shubham
Many enterprise projects now pass free text to large language models (LLMs), and that text often contains personally identifiable information (PII) such as names, phone numbers, email addresses, home addresses and dates of birth. Once a prompt is sent, this data is copied into provider systems, application logs and observability traces. Redacting PII before the call is an obvious control, yet teams seldom have numbers on what it costs in latency or what it lets through. This paper compares four redaction designs that any LLM pipeline can adopt: a pure regular-expression (regex) redactor, Microsoft Presidio running on a spaCy model, a hybrid in which Presidio finds only person names while regex handles structured identifiers, and a fast hybrid that calls spaCy's named-entity component directly. All four share a known-entity rule, which removes the person the system already knows it is handling, and a consistent placeholder scheme, so the LLM can still tell one person from another after redaction. We evaluate the designs on a sample workflow, which we call ProfileAssist, that turns resumes into structured profiles, using 100 synthetic resumes (523 characters on average, with 900 PII values and 600 business terms that must survive). Regex ran at a 0.20 ms median but caught only 88.9% of PII, because it cannot find names that are not known in advance, such as the references listed in a resume. Presidio, hybrid and fast hybrid each caught 100% of PII and kept 100% of business terms, at median latencies of 37.59 ms, 30.23 ms and 13.48 ms. Measured against a 9,120 ms median LLM call, even the slowest option adds only 0.41% to the request. We therefore recommend the fast hybrid as a default for LLM pipelines that handle personal data, with regex kept as a safety net rather than as the only control.
Ke Cheng, Xin Xu, Zhichao Chen, Yixiao Chen, Lei Xin, Jianbo Zhao, Fanhu Zeng, Yue Liu 等 10 位作者
Induction circuits retrieve a value by matching a query to information carried from the token before that value. How this address forms, incorporates context, and supports sequential retrieval remains less clear. We use hybrid language models as a tractable testbed and identify the historical source key as a common causal interface between local preparation and global retrieval. Probes and restoration interventions connect this interface across controlled models, pretrained recurrent and sliding-window hybrids, and released language models. We organize the analysis by address context radius and sequential retrieval depth. First, shared local parameters couple predecessor Carrying to recurrent state processing; separating these demands or biasing initialization promotes address formation. Second, adapting input representations extends an established literal circuit to category matching, making semantically compatible records active competitors. Earlier context resolves this competition through the source key, with a sharp visibility boundary in a controlled single-read model and compositional relays across local layers. Third, fixed relations support memorization shortcuts, whereas sequential generalization requires learning an output-to-query bridge between retrievals. Controlled state-tracking and compositional in-context tasks demonstrate this requirement across stacked stages, looped recurrence, and intermediate-token generation. Together, these results connect circuit formation, contextual disambiguation, and sequential use through a shared retrieval interface. Code is available at https://github.com/ckpassenger/bind-match-copy.
Xiaodong Wu, Xiangman Li, Jianbing Ni, Rongxing Lu
An image editing service records which user requests each edit, but the edited image itself does not contain this account information. Once the image is downloaded and shared, the service cannot identify the requesting user from the image alone. Embedding a user-specific watermark could enable such attribution, but semantic editing may erase the signal. We propose Tracemark, an editor-aware watermarking framework for identifying users from edited images. A shared encoder embeds a user-specific signal before editing, and a shared detector recovers the user by comparing features of the edited image with learned reference vectors. Both components are trained jointly through an editor with fixed parameters. User identification errors encourage the encoder to produce watermarks that survive editing, while fidelity losses limit changes to the input image and intended edit. Each image domain and editing instruction uses a separately trained model shared across registered users. Experiments on three image domains show that editor-aware training improves user recovery with both reference-vector matching and binary watermark decoding. Separately trained models reach 87.5% mean identification accuracy among 500 users. When tested with a different editor without retraining, the watermark modules still recover user information. These results demonstrate that learning watermarks through the editing process enables user attribution without modifying the editor.
Mayank Anand, Subham Raj, Aman Vaibhav Jha, Sriparna Saha
Long-horizon LLM agents require memory systems that preserve useful information across sessions and retrieve it when needed. Existing systems treat these as separate problems: memories are written prospectively with static importance scores that are never revised after use. This creates two coupled failure modes: consolidation can silently destroy evidence that future queries require, while retrieval continues to favor stale memories. We propose MEMOIR (Memory Evidence via Merge-certified, Online Importance Revision), a training-free memory lifecycle that unifies judge-gated consolidation with retrospective attribution. Before compressing memory, screening gates test whether answer-relevant evidence is preserved. After each query, attribution estimates which retrieved memories supported the answer and updates their retrieval priority. A UCB-style exploration bonus additionally encourages retrieval of under-exposed memories, reducing the risk that frequently retrieved memories crowd out useful but less-seen evidence. On 1,905 deduplicated LoCoMo QA pairs, the full retrieval stack improves evidence recall@5 by 11.92 percentage points over bare cosine, with confidence-gated feedback adding 1.50 points overall without reference answers or reranker training; a deployment-unavailable oracle adds 1.84 points overall and 8.05 points on multi-hop. Judge-gated merge changes the recall of text-based evidence from 58.82% to 59.08%, reduces the active store by 1.42%, and incurs 0 observed losses in 100 independently audited merges, with a one-sided 95% upper loss-rate bound of 2.95%. Together, these results outline a practical deployment path: no additional LLM call on the online write path, with merge screening confined to periodic offline consolidation. The lifecycle reduces to one principle: attribute usefulness after use, and screen for preservation before compression.
Event cameras are neuromorphic vision sensors with pixels that independently and asynchronously respond to brightness changes at microsecond resolution, supporting state estimation under rapid motion, motion blur, and high dynamic range (HDR) illumination. This also enables capture of high-frequency blinking patterns from light-emitting diodes (LEDs). Based on this, we propose a real-time event-based framework for LED tracking and six-degree-of-freedom (6-DoF) relative pose estimation using pulse-width modulation (PWM)-encoded LED markers. Multiple LEDs share a common blinking frequency and encode identities through distinct PWM duty cycles, reducing frequency-resource requirements and computational cost. Our dual-rate asynchronous tracking framework combines low-rate LED detection with high-rate raw-event updates, maintaining robust LED identity association and accurate pixel-level tracking under rapid motion. LED identities and positions establish reliable 2D-3D correspondences for real-time metric 6-DoF relative pose estimation. Extensive evaluations on a self-collected dataset demonstrate robustness under large motion flow and transient image-plane LED overlap. Furthermore, we develop a coordinated robotic platform comprising a quadrotor and a Mecanum-wheeled mobile robot, demonstrating the proposed framework's potential for closed-loop drone guidance through hovering and target-following experiments under time-varying HDR illumination, with onboard deployment on an NVIDIA Jetson Orin NX. The dataset and implementation will be released to facilitate further research.
We study adversarial online learning with expert recommendations in a totally ordered action space, where the experts' recommendations may change and be reordered arbitrarily. Our focus is two-point feedback: on each round the learner plays two current recommendations, observes both, and pays their average. We study two problems. In the separated monotone problem, the loss is the sum of a hidden nondecreasing component and a hidden nonincreasing component, and a query reveals the two components separately. In the BV problem, the losses are bounded and have bounded variation (BV) along the order, and a query reveals only the loss. In the separated monotone problem, two paid observations achieve \[ O\bigl(\log(|\cF|T)\sqrt{T\log|\cF|}\bigr) \] regret against the best of $|\cF|$ experts, within a logarithmic factor of the classical full-information bound. For the BV problem, we obtain $\wtO(\sqrt T+|\cF|)$ regret. We give an application in dynamic pricing with expert advice, showing that two-option menus matches the full information rate up to $\log$ factors. The main technical ingredient is a marginal-preserving endpoint flow estimator. Two observations reveal a loss increment along the order; random paths supply telescoping identities that turn this local information into unbiased loss estimates for all current recommendations. A reversible pair sampler preserves the learner's intended marginal at both played actions. The estimator's second-moment and stability bounds support online updates for both feedback models.