Streaming video understanding requires models to repeatedly decide whether to respond or continue observing. Recently, Group Relative Policy Optimization (GRPO) has been applied to this setting, with existing approaches emphasizing reward design while retaining a single trajectory-level advantage across turns. Such shared credit can reinforce redundant replies after earlier successes and discourage useful follow-ups after earlier failures. We propose TurnWise, a turn-level GRPO framework that aligns learning with individual streaming decisions. TurnWise addresses the mismatch between trajectory-level feedback and turn-level interaction, promoting informative responses and reducing redundancy. Specifically, turn-level credit assignment uses silence-completed prefix scoring to construct group-relative advantages from each decision's remaining return, separating its learning signal from earlier rewards. Turn-mean policy optimization averages token losses within each turn and then across turns, giving response and waiting decisions equal weight regardless of output length. These complementary designs improve streaming policy learning without a learned critic or additional rollouts. Extensive experiments show our method outperforms prior RL approaches for streaming video understanding.
Recommender systems usually learn what people want from what they have done. Platforms could also ask people directly, but asking is costly: 98.8% of the prompts in our data went unanswered. Theories of statistical prediction imply that a person's own forecast adds little to a recommender algorithm built from many people's behavior. Theories of dispersed knowledge imply that people hold information about their own taste that the data lack. We test the two with the MovieLens Beliefs dataset. In 2023-2024, a homepage prompt asked members how they expected to rate movies they said they had not seen. Of 26,020 such expectations, 551 from 277 members were followed by a rating at least a day later. We compare each expectation with four benchmarks: the platform's own predicted rating, two collaborative-filtering models (matrix factorization and an item neighborhood), and a stack trained on other members' outcomes. The trained models use each member's latest ratings. People predict their own ratings more accurately than all four benchmarks (correlation 0.69, against 0.34 for MovieLens and 0.54 to 0.59 for the trained recommenders). Added to all the machine forecasts and features, the expectation lowers out-of-sample mean squared error by 26%. The gain holds when models are trained on earlier expectations and tested on later ones. A member's earlier expectations about other movies do not replace the current one. In ordering a member's own movies, the trained recommenders are statistically indistinguishable from the person. On old, heavily rated movies, the machines do as well or better. The results reconcile the two theories: the crowd's data suffice where they are rich, and the person adds information about the specific item. We derive four design principles for preference elicitation and recommender evaluation.
Jiahui Sheng, Le Li, Limin Lin, Jiahao Li, Wansen Feng, Hong Gu, Shuhan Chen, Xiaorun Li 等 9 位作者
Portrait video relighting aims to modify the illumination of a portrait video while preserving photorealism and temporal stability. The task becomes particularly challenging under dynamic lighting, where illumination direction, intensity, and color vary over time. To tackle this problem, we introduce FlashRelight, a novel two-stage framework for dynamic portrait video relighting. FlashRelight first suppresses the source illumination through delighting, and then synthesizes the target dynamic lighting on an illumination-neutral portrait video. FlashRelight supports three complementary control modalities: envmap sequences for explicit frame-level lighting control, dynamic lighting text for semantic illumination editing, and lighting reference videos for visual lighting transfer. Training follows a low-cost hybrid strategy that combines static one-light-at-a-time (OLAT) data with generated dynamic-lighting portrait videos, covering both controlled illumination variation and realistic portrait motion. Extensive experiments show that FlashRelight achieves temporally coherent and photorealistic dynamic relighting while enabling flexible cross-video lighting transfer. Code is available at https://github.com/vivoCameraResearch/FlashRelight
KV cache compression reduces the communication volume of disaggregated large language model serving, but its practical benefit depends not only on compression ratio and model quality, but also on how efficiently the compression pipeline executes. This requirement becomes increasingly important as modern serving systems adopt high-bandwidth interconnects, where inefficient compression can offset a substantial portion of the saved communication time. This paper presents FastKVComp, a high-throughput KV cache compression data path built on top of KVServe. FastKVComp improves compression execution in two complementary ways. First, it fuses the distribution transformation and low-bit quantization stages into a KVaware TileLang kernel, eliminating intermediate tensor materialization and redundant memory accesses. Second, it replaces the original lossless compression backend with an open, extensible, and searchable backend based on LC, enabling high-performance compression operator pipelines. The fused kernel achieves up to 5.9x speedup, while the new lossless backend increases compression throughput from below 90 GB/s to over 500 GB/s. Across the evaluated workloads, FastKVComp reduces job completion time by up to 2.04x. End-to-end evaluation shows that these optimizations allow KV compression to retain meaningful communication benefits under modern high-bandwidth network.
Variational approaches to learning transport maps for multimodal sampling can suffer from mode collapse. One way to mitigate this problem is to learn transport along a prescribed path of intermediate distributions defined by geometric annealing. We show that annealing can take a substantial detour in Wasserstein space: when mass moves between modes, the path can be much longer than the distance between its endpoints. We further link path geometry to transport learning, showing that long annealing paths can slow population optimization, while sufficiently regular paths permit rapid convergence. Motivated by these findings, we propose componentwise annealed transport (CAT) to reduce the need for mass transfer between modes. Starting from a mode-informed mixture reference, CAT constructs an exact mixture decomposition of the target and learns separate transport maps along componentwise annealing paths. We estimate the component weights separately and show that these estimates remain consistent even when the learned flows are inexact. We also develop adaptive annealing schedules guided by transport effort. Experiments on multimodal synthetic distributions and Bayesian inference problems demonstrate improvements in transport learning and sampling efficiency.
LLM agents are judged by what they deliver, and in professional work what gets delivered is a project, not a task. A project pays off only if the whole package holds, and one weak part can limit the value of the rest. Yet existing benchmarks mostly test agents on one task at a time, leaving a gap between the capability we measure and the outcome that makes agents useful. To address this gap, we introduce DeliverBench, which tests whether agents can carry a project from brief to final delivery. Its 11 projects, spanning financial due diligence, legal contract drafting, and drug safety reporting, are reconstructed with 13 domain experts from real materials and chain 150 subtasks through 2-15 rounds of new materials and client feedback. Because no reference outputs are injected, agents inherit their own errors from one task to the next. Expert rubrics then score every checkpoint as well as the final package. Across seven model-harness systems, even the best system scores only 65.0 out of 100 on final delivery and passes 50.0% of subtasks at a 90-point threshold. The failures are often failures of carrying work forward: in one audit project, all seven systems addressed the review findings, but only one preserved traceability. Together, these results suggest that task-level capability does not simply compose into delivery, which also depends on keeping the work coherent as it evolves.
Scientific discovery depends on reusing experimental knowledge, yet records scattered across papers and repositories are often incomplete. In materials synthesis, recovering missing conditions requires reasoning about the target material and the observed context of a specific experiment. Hyper-relational knowledge graphs (HKGs) represent this context through qualifiers, but existing completion methods leave domain knowledge and experimental regularities implicit. We propose RAHC, a rule-augmented HKG framework for single-slot and joint multi-slot synthesis-condition completion. RAHC builds a traceable rule catalog from training records and retrieved literature, retrieves rules for each query-candidate pair, and separately encodes graph structure and rule evidence. Hard rules constrain candidate validity, while soft rules guide ranking. RAHC-G aggregates rules using global reliability; RAHC-Q additionally learns query-specific relevance. Gated fusion combines structural and rule representations, while tuple-conclusion rules provide cross-condition evidence for joint completion. Experiments on five materials-synthesis datasets show improvements in single-slot and joint completion under Semi-inductive and Open-Answer Inductive settings, with Semi-inductive MRR gains of up to 15.6% over MAYPL. RAHC also improves performance on the entity-disjoint ILPC'22 Small benchmark, supporting transfer to fully inductive KG completion. These results support experimental condition selection, synthesis planning, and the reuse of scientific knowledge.
Large language models (LLMs) are increasingly expected to extract and synthesize structured information from complex real-world documents, yet they typically produce unstructured, hard-to-verify paragraph-style outputs. We introduce AOE (Arranged and Organized Extraction), a bilingual benchmark that evaluates LLMs' ability to distill long-form, multi-document inputs (64k to 1.8M tokens) into structured tables. AOE comprises 12 tasks across four domains (academic, financial, legal, and medical), each grounded in genuine analytical needs, with schema guidance provided to isolate extraction quality from formatting variance. We evaluate 12 frontier LLMs (both open-weight and proprietary) under four generation settings of increasing autonomy: direct full-document input, RAG-based retrieval, offline document-browsing agent, and online web-search agent. Even the best-performing model achieves only 42.7% Cell F1, and retrieval-augmented approaches often degrade rather than improve performance, revealing that structured extraction from long contexts remains a fundamental challenge - an "Area of Effect" (AOE) that impacts all tested systems.
Long Zhang, Yuhan Chen, Chaoran Zhang, Wanxia Cao, Kun Huang, Pengzhi Gao, Wei Liu, Jian Luan 等 10 位作者
Vision-Language Models (VLMs) based GUI agents stand to benefit significantly from online reinforcement learning (RL). However, their training is bottlenecked by two fundamental issues: current data synthesis methods for GUI Agents rely on specific environments and struggle to generate diverse data, while existing evaluators either suffer from limited scalability or provide inaccurate and unreliable reward signals. To overcome these challenges, we introduce GSAR (Goal-State-Anchor Reward), a RL reward framework that supports scalable task generation and delivers reliable reward signals for stable and efficient policy optimization. Our approach features self-evolving data synthesis, which produces multiple environments through task execution and generates diverse tasks and goal states. Complementing this, a state-anchor mechanism automatically annotates task-relevant UI elements in successful goal states as reference anchors. During RL training, these reference anchors provide accurate, scalable reward signals that substantially enhance efficiency. Extensive evaluations demonstrate that our framework achieves over 90% accuracy on offline trajectory verification and performs closest to rule-based methods. Furthermore, agents trained using our reward framework exhibit strong performance on both AndroidWorld and our constructed benchmark, establishing a scalable approach for GUI agent training.
Novel view rendering of large and complex reconstructed scenes is becoming increasingly photorealistic. However, most reconstructions remain static and lack the ambient motion that makes environments immersive. We present AniGS, a method for scene-level animation of 3D Gaussian Splatting (3DGS) reconstructions that adds subtle, distributed dynamics, e.g., vegetation motion, while preserving rigid structures. Unlike existing 3D animation techniques which are limited to object-centric subjects or small regions, AniGS is designed for large, cluttered, navigable scenes. AniGS represents the scene with a canonical 3DGS and models motion using a time-conditioned deformation field. To animate the entire scene, we leverage a pretrained video diffusion model and introduce an iterative dataset--model update strategy that progressively expands viewpoint coverage and repeatedly updates camera-fixed training videos using a render-and-refine scheme. To prevent artifacts from unintended motion in static areas, we further introduce a composed video-to-video refinement scheme that restricts motion to desired regions. Experiments on five real-world, large-scale outdoor scenes demonstrate that AniGS produces natural ambient dynamics and high-quality novel view videos, enabling more immersive viewing experiences of reconstructed environments.
Yihan Chen, Jialin He, Xihe Li, Jie Ma, Tianying Xie
Given a graph $T$ and a nonempty family $\mathcal{F}$ of graphs, let $\operatorname{ex}(G, T, \mathcal{F})$ denote the maximum number of copies of $T$ in an $\mathcal{F}$-free subgraph of a host graph $G$. Alon and Shikhelman [Algorithmica, 2022] conjectured that if $e(T)> 0$ and no member of $\mathcal{F}$ is a subgraph of a blow-up of $T$, then for every $\varepsilon>0$, approximating $\operatorname{ex}(G, T, \mathcal{F})$ within an additive error of $n^{v(T)-\varepsilon}$ is NP-hard for an input graph $G$ on $n$ vertices. Previous advances toward this conjecture include the resolution of the canonical case $T=K_2$ by Alon, Shapira, and Sudakov [Ann. of Math., 2009], as well as the case of $\operatorname{ex}(G,K_m,K_k)$ by Alon and Shikhelman. In this paper, we confirm this conjecture for every bipartite graph $T$ and every clique $T$. Our proof reductions combine several techniques, including iterative leaf additions, symmetric rooted extensions, and random transversals. Together, they encode the optimal value of the NP-hard Max $r$-Colorable Subgraph problem into the objectives $\operatorname{ex}(G, T, \mathcal{F})$ for a suitably constructed graph $G$.
Platforms ask customers for online reviews to reduce self-selection, but theory conflicts on whether review solicitation makes online ratings more or less informative. Selection theory implies that asking helps, voice theory that it dilutes the news, and research on rating inflation that compression destroys information. We reconcile these views with signal detection theory, which separates what a rating knows (its sensitivity) from what it shows (its criterion). We test the framework on Steam, a leading PC video game store, whose client began in October 2019 to prompt players for a review after about three hours of play. Using a difference-in-discontinuities design at that inferred threshold in 21.6 million Steam reviews of 315 games, we find that review counts just above the threshold become 4.25 times those just below, and the jump in the positive share grows by 6.6 percentage points. Agreement with critics changes little near launch and erodes later. The rating's movement changes far more: pass-through from reviews written in the first two hours falls from 0.73 to 0.18 just above the threshold, against 0.81 to 0.64 just below, and a fixed rating tier responds to 12% of downturns in those reviews above the threshold, down from 42%, against 41% and 47% below it. What the stream knows changes less. At equal review counts its sensitivity falls more above the threshold, because added reviews respond about half as much and scores near the ceiling are noisier per review, but at natural counts added volume offsets this loss, leaving a change of -0.03 relative to the stream below (95% CI -0.13 to 0.06). Failures that developers acknowledge, which run through complaints, still move the score after launch. Review prompts therefore mainly change how a rating must be read, and platforms that solicit reviews should recalibrate their rating labels.
Safe autonomous driving requires more than recognizing the current scene: the ego vehicle must anticipate how surrounding traffic participants will evolve under its own motion. However, existing end-to-end driving models typically encode scenes into dense visual or BEV features, leaving object-level interactions implicit. We propose (INTERACT INTERaction-aware Action-Conditioned Trajectory learning), a three-stage framework that learns a slot-based object-centric world model from unlabeled driving videos and transfers it to trajectory planning. INTERACT first learns recurrent slots that group dense visual features into temporally structured scene-entity representations without object annotations. It then infers latent ego actions from slot context and video transitions, and uses them to condition future slot prediction, capturing how ego motion and surrounding entities co-evolve. Finally, the learned slots, latent actions, and future slot predictions are distilled into a trajectory planner to guide trajectory generation and selection. Extensive experiments on NAVSIM and nuScenes validate the effectiveness of INTERACT. In particular, it achieves state-of-the-art performance on NAVSIM, with improvements in safety-critical collision metrics.
Yaochen Wang, Chenguang Wang, Ming Li, Haomin Zhuang, Xinyue Zeng, Han Chen, Kwesi Adu Cobbina, Dang Nguyen 等 11 位作者
AI reviewers are increasingly used to assess scientific manuscripts, yet it remains unclear whether their judgments reflect the improvements authors make through revision. We investigate this question using PaperTrajectory-2026, the 2026 release of PaperTrajectory, a dataset of 6,959 paper trajectories across five AI conferences, with automatic annotations of changes along 14 dimensions and separate judgments of overall quality. The annotations show that major improvements emphasize empirical evidence and framing, while later revisions tend toward incremental improvement. Using independent assessments of manuscript versions from eight AI reviewers, we examine whether annotated major improvements receive higher scores, which revision dimensions are associated with score changes, and whether user opinions can outweigh observed revision gains. Every reviewer leaves scores unchanged on most major improvements, and agreement on the direction of score changes is weak. Revision dimensions show some reviewer-specific associations with score increases, but provide limited predictive discrimination and no consistent guide across reviewers. Even when a revision earns a higher neutral score, a brief negative user opinion can erase or reverse that advantage, with substantial variation across reviewers. These findings show that substantive revisions are not consistently reflected in review scores and that observed scoring gains can be vulnerable to user framing, motivating assessments that connect manuscript evidence to overall judgments and remain stable when no scientific evidence changes.
Crowds Believe in You: Online Prediction Markets Overestimate How Often People Achieve Their Own Goals
Zihan Chen, Lei Nico Zheng, Di Zhu
AI 速览：基于 Manifold 上 6,150 个自我目标预测市场的订单重建，作者报告自我市场比外部事件市场更常被高估，且纠偏信息主要掌握在当事人手中。
People increasingly share personal goals online and invite others to forecast them. Research on the planning fallacy finds that observers predict a person's behavior less optimistically than the person does, and prediction markets aggregate such judgments, harnessing the wisdom of crowds. We study 6,150 prediction markets on Manifold in which 1,660 users asked about their own futures ("Will I ...?"), reconstruct every order and execution, and compare these self-markets with markets about external events. Self-markets resolved YES 4.4 percentage points less often than their final prices implied, whereas external markets were nearly calibrated (0.9 points); the difference was 2.2 points at the same price level and 1.9 points within creators who made both kinds of market. Absent when markets opened, the overstatement emerged within a day of trading, even where the person never traded, and peaked in the week before close. In a market's first day, observers moved external-market prices down toward the base rate, but moved self-market prices 3.6 points further toward YES, except in markets naming an outcome the person would not want, which were not overpriced. The correction came late and largely from the person, whose orders supplied 38% of the eventual movement toward the outcome in self-markets, against 9% in external markets. Because self-market prices mainly failed to separate successes from failures, a recalibration learned from earlier markets improved later forecasts by only 1.5%. Crowds believe in the futures people pose for themselves, and the information that would correct them lies with the person.
AI Cosmos Team, Jinfeng Zhou, Yuxuan Chen, Fangming Chang, Mingyu Su, Yifan Zhu, Yuan Wang, Bo Zhang 等 35 位作者
Large language models (LLMs) can improve task performance during learning, yet these gains do not necessarily reflect learners' ability to thereafter apply the knowledge and skills independently. We present AI Cosmos, an agentic infrastructure for situated learning that organizes learning around two complementary principles: scenario-based participation and cumulative assets of practice. AI Cosmos organizes learners and role-specialized agents within learning scenarios that integrate task-relevant knowledge and, when needed, executable workbenches. An asset-memory loop preserves the processes and outcomes of practice as reusable assets with explicit provenance and dependencies. We evaluate scenario-based interaction through two preregistered randomized controlled studies examining the acquisition of legal reasoning and programming competence, holding the same underlying LLM across conditions. AI Cosmos achieves greater learning gains than a conventional LLM in legal reasoning and higher mastery of unseen programming problems than an LLM-as-Teacher baseline.
Reinforcement learning has advanced large language models on tasks with reliable verifiers, where checking an outcome provides a clear basis for reward. Extending this approach to open-ended tasks requires environments that can both evaluate complex, partially successful outputs and turn these judgments into effective learning guidance. We propose a reinforcement learning framework built around a training environment that integrates human domain knowledge with model reasoning. Human domain knowledge, encoded in rubrics, anchors evaluation to task intent, while model reasoning combined with hierarchical rewards distinguishes complex open-ended outputs and links diagnosed failures to likely responsible output regions. Furthermore, these diagnostic attributions guide token-level weights in the GRPO objective, thereby translating the ambiguous notion of success in open-ended tasks into concrete guidance for policy optimization. We instantiate this framework in interactive web application generation, a representative open-ended generation task. The resulting model achieves an average pass rate of 41.25\% on MiniAppBench and 76.19\% on ArtifactsBench. Ablation studies show the contributions of individual components, while additional out-of-distribution evaluations support transferable gains without degrading performance on other evaluated tasks.
Ajay Vikram Periasami, Anjaneya Gupta, Ethan Cheng, Ivan Rosales
Balatro, a 2024 roguelike deckbuilder, presents three difficulties that have historically derailed reinforcement learning methods: a 3,000-action state-dependent action space, horizons of several hundred decisions with sparse rewards, and pervasive stochasticity in both card draws and shop generation. Published RL results on the game remain scarce. We introduce an end-to-end pipeline comprising (i) a high-throughput Rust simulator that is automatically state-diffed against the live game client via BalatroBot, (ii) a condensed observation-action interface with dynamic per-step masking of legal moves, and (iii) a multi-agent PPO architecture that decomposes play into a tactical blind policy and a strategic shop policy that share the full game state but act in disjoint phases. Long-horizon credit assignment is partially addressed by per-agent reward routing with explicit cross-agent credit. On the main joint run, training metrics show a sharp phase transition in which the first-blind clear rate converges above 99% and mean episode return increases by roughly an order of magnitude, plateauing at a mean maximum ante of 3 out of 8; single-agent baselines under identical rewards oscillated between greedy and hoarding failure modes and did not stably clear the first blind. These results establish the two-policy decomposition as a strong reactive baseline for Balatro and localise the remaining gap to long-horizon strategic planning rather than tactical blind play.
Distributional metrics such as Fréchet Audio Distance cannot score individual music clips and correlate poorly with human judgments, while the only per-sample learned metric achieving high human correlation is closed-source. We introduce MUQ-EVAL, an open-source per-sample quality metric for AIgenerated music built by training lightweight prediction heads on frozen MuQ-310M features using MusicEval, a dataset of generated clips from 31 text-to-music systems with expert quality ratings. Our simplest model, frozen features with attention pooling and a two-layer MLP, achieves system-level SRCC = 0.957 and utterance-level SRCC = 0.838 with human mean opinion scores. A systematic ablation over training objectives and adaptation strategies shows that no addition meaningfully improves the frozen baseline, indicating that frozen MuQ representations already capture quality-relevant information. Encoder choice is the dominant design factor, outweighing all architectural and training decisions. LoRA-adapted models trained on as few as 150 clips already achieve usable correlation, enabling personalized quality evaluators from individual listener annotations. A controlled degradation analysis reveals selective sensitivity to signal-level artifacts but insensitivity to musical-structural distortions. Our metric, MUQ-EVAL, is fully open-source, outperforms existing open per-sample metrics, and runs in real time on a single consumer GPU. Code, model weights, and evaluation scripts are available at https://github.com/dgtql/MuQ-Eval.
Generative AI now produces images and poems that people cannot reliably distinguish from human-made work, so people often judge for themselves who made a work. We examine how these perceived-authorship judgments relate to aesthetic preference. In a public online AI Art Turing Test, 11,123 people judged whether each of 50 pictures (26 AI-generated) was made by a human and were asked for their favourite. Among the 9,152 respondents who judged their favourite, a picture was 22.1 percentage points more likely to be judged human-made when it was the respondent's favourite, holding respondent and picture fixed. The association was similar for AI-generated and human-made favourites, was positive for all 50 pictures, and appeared under every reported detection strategy. Respondents most negative about AI art chose fewer AI-generated favourites (47% against 60% among the most positive) but more often believed them human-made, partly because others also judged those pictures human. As a result, in every opinion group a third to two fifths of respondents named an AI-generated favourite they believed human-made. A reanalysis of a published experiment on AI-generated poetry, in which 230 people rated each poem before judging its authorship, found the same pattern, although the AI poems were rated higher. When provenance is hidden, liking and perceived human authorship travel together, consistent with aesthetic response serving as evidence of a human maker and with beliefs about the maker shaping enjoyment. The findings qualify research on bias against AI art, which typically uses explicit labels, and inform AI disclosure policy.
Attack logs and red-team reports show targets the attacks used against them, to help them repair their defenses. Prompt injections against large language models (LLMs) are plain text, so such records also hand targets a working copy. Drawing on social learning theory and work on the victim-offender overlap, we study what targets do with that copy in Tensor Trust, an online game whose players defended LLM-guarded accounts, attacked one another's, and could read every attack they received. Using all 563,349 attack records, we compare each received attack with matched attacks that circulated among similar players at the same moment but never reached the target. Reuse mainly followed success: within a week, targets attacked other players with 10.0% of the attacks that had broken their defense, versus 0.03% of matched attacks (adjusted difference: 8.8-10.0 percentage points). Near-misses, whose replies began with "access granted" but that scored as failures, were reused about as rarely as failures: reuse tracked whether an attack paid off, not how its reply read, consistent with payoff-biased social learning. Breaches were also followed by defense changes, after which same-family attacks became 24.5 percentage points less successful against the target, relative to other defenders. Reuse came mainly from players already attacking, not newcomers. Previously received texts accounted for 23.2% of the successful hijacks we analyze. The attack history thus served both repair and reuse; we derive testable hypotheses for how AI security tools should disclose attacks, such as describing the technique without the full text.
Under a fixed embedding-storage budget, recommenders can use a small full-precision model, a distilled student, or a higher-rank low-precision model. We asked which choice preserves rare-item recall and whether teacher fidelity reflects it in seven dataset-backbone cells: matrix factorization on Gowalla, Last.fm, Amazon Video Games, Steam, Goodbooks-10k and MovieLens-20M, and LightGCN on Amazon Video Games. Seven candidates fit in 168 bytes per user or item row: 40-dimensional Bayesian personalized ranking (BPR), two 40-dimensional unbiased knowledge distillation (UnKD) students, and factorizations of a 300-dimensional teacher's scores at ranks 40, 80, 155 and 300, stored in 32, 16, 8 and 4 bits. Each test set was evaluated once after fixing candidates and confirmatory analyses. The rank-300 4-bit factorization used 13% of teacher storage and retained 95-103% of its tail recall@20 and 93-97% of its NDCG@20 in every cell. The rank-40 factorization had the lowest tail recall in every cell; the 4-bit factorization had 1.2-12.7 times its tail recall and 0.99-1.46 times its NDCG. UnKD students retained 56-102% of teacher tail recall. In all ten prespecified comparisons across five prospective-expansion cells, the 4-bit factorization had higher tail recall, with simultaneous 95% intervals excluding zero. Teacher-hit retention, the share of the teacher's relevant tail hits also found by a model, ordered 23 of 147 candidate pairs differently from tail recall and identified the wrong tail-recall leader in 3 of 7 cells: models can find relevant tail items the teacher misses. In these cells, spending bytes on rank rather than precision preserved rare-item recall at 7.5 times the arithmetic per score. Compressed recommenders should be judged by total tail recall, not teacher fidelity alone.
Diffusion-based real-world image super-resolution suffers from limited reconstruction fidelity due to information loss in VAE-based latent representations. Recently, pixel-space diffusion models offer a promising high-faithfulness alternative by avoiding latent compression, but adapting them to Real-ISR remains challenging due to inherent patch boundary artifacts and degraded generative capability. To address the issues and maximize the potential of pretrained pixel-space diffusion models, we propose HiFiPix, a pixel-space Real-ISR framework capable of ultra high-fidelity reconstruction and fine-grained visual generation effects. To suppress patch boundary artifacts, we penalize unrealistic high-frequency periodic components with a Cross-Scale Wavelet Discriminator (CWSD) for image-level supervision, and Mixup Contrastive Self-Distillation (MixCSD) for intermediate representation regularization. Furthermore, we uncover a heterogeneous generation-fidelity trade-off across low-rank adapters (LoRA) and propose Selective LoRA Pruning (LoRA-SP) to adaptively prune LoRA layers, enhancing generative capability with minimal deterioration of reconstruction fidelity. Extensive experiments demonstrate that HiFiPix consistently outperforms existing state-of-the-art latent space methods in terms of both restoration accuracy and perceptual quality by a large margin. Code and models will be released.
Aman Tyagi, Hemanth Boinpally, Jonathan Chen, Douglas Gebert, Steven Hickson
Modern black-box Image-to-Video (I2V) models offer powerful capabilities in automated content creation, yet their lack of fine-grained control and reliability presents significant challenges in professional workflows. Their inherent stochasticity causes minor variations in textual prompts or hyperparameters to yield drastically different outputs often necessitating inefficient, brute-force trial-and-error processes. To address these limitations, we introduce the ``Agentic Self-Improvement" framework, which reframes video synthesis into a closed-loop, goal-directed optimization. Our framework systematically navigates the generation parameter space using a novel two-stage approach. In the first stage, an iterative prompt optimization loop uses a multimodal Large Language Model (mLLM) to refine the input prompt. This refinement implements two automated evaluations: Davidsonian Scene Graph (DSG) queries ensure semantic adherence, and Common Mistake Questions (CMQ) for artifact detection. At the second stage, we use Bayesian optimization to efficiently co-optimize stochastic seeds and CFG scales. This search is guided by a suite of quality metrics, including the novel Video-Text Adherence (VTA) score derived from the DSG and CMQ evaluations. Our framework significantly outperforms unguided search methods: in human preference studies, videos generated via our agentic approach were strongly preferred over baseline outputs, achieving win rates up to 69\%. This work provides a practical and extensible methodology for enhancing the predictability and control of state-of-the-art video generation models, moving the field beyond speculative curiosities toward reliable, production-ready tools.
Hanze Li, Athan Z. Li, Feiyang Kang, Jiachen T. Wang, Yuhan Chen, Ruishan Liu, Ruoxi Jia
Data curation is a first-order design choice in machine learning, yet tailoring training data to a downstream model and task remains costly in labor, expertise, and compute. Datasets are therefore often processed through fixed pipelines and reused across applications. We show that this assumption is fragile: across 55 dataset--model tasks spanning biological domains, the same curation recipe can help one model while hurting another. Effective curation must adapt to the downstream setting, motivating automated search over data recipes. We first build an agentic system that searches for curation policies at deployment, improving data utility via \textit{test-time curation (TTC)}. Although achieving performance gains, its search is driven largely by recent feedback and remains local. Adding persistent memory helps the agent retain evidence across iterations and reveals a deeper problem: it forms claims that extend beyond the conditions in which they were observed, sometimes abandoning an entire strategy after a single unsuccessful trial. To address this problem, we introduce CaliMem, a calibrated-memory agentic loop that requires the agent to predict the outcome of each proposed policy before executing it. Comparing this prediction with the observed result gives the agent a \textit{reality check} on the claims behind its proposal. It uses the prediction error to retain supported claims, revise mistaken ones, and specify the conditions under which a conclusion holds. This makes the agent's beliefs testable, reduces overgeneralization, and supports more effective long-horizon exploration. Across biological and LLM post-training tasks, CaliMem consistently outperforms fixed curation baselines and published methods, and exceeds full-data training by up to 19.4\%. Code is open-sourced.
We deploy large language models (LLMs) as business development (BD) agents for persuasive price negotiation in online travel agencies (OTAs). The agent must follow a multi-stage Standard Operating Procedure (SOP) and strict guardrails (no over-promising and no hallucinations), while remaining human-like and effective over long, multi-turn dialogues.We propose Reward-Enhanced Policy Optimization (REPO), a reinforcement learning post-training method that combines heterogeneous rewards: a preference-trained reward model (RM), an LLM-as-a-judge (RJ) for nuanced behaviors (e.g., emotional value and SOP compliance), and rule-based reward functions (RF) (mainly regex-based) for deterministic checks on numerics, formatting, and guardrails. In expert consensus evaluation (three human experts; 30 online conversations and 45 curated bad cases), REPO improves average dialogue rating to 4.63 (+0.33 over GRPO) and raises the share of conversations with at least one excellent response to 66.67% (+23.34 pp over GRPO), while achieving a 93.33% bad-case fix rate with 75.56% clean fixes.In a production A/B test on 9,653 real customer conversations (vs. an intent-driven dialogue system), REPO improves response rate by +12.14 pp and task success rate by +5.94 pp (p<0.001).
尚无 AI 解读，展示原文摘要
cs.CL
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 6: Industry Track) · 2026 · 正式发表
Yuchong Chen, Bowei Zou A*STAR, Yuhan Chen Soochow University, Yifan Fan, Xinyu Li, Shujun Cao, Yu Hong Suzhou University
Visual questions are often ambiguous: the same image–question pair may admit multiple valid answers depending on which region is referenced. However, current Visual Question Answering (VQA) systems typically collapse this ambiguity, committing to a single interpretation during decoding and evaluation. In this work, we study visual question ambiguity from a grounded, region-centric perspective. We operationalize ambiguity as the existence of multiple distinct answer-supporting regions in an image, each independently yielding a valid answer. This formulation makes ambiguity observable without requiring exhaustive multi-answer annotations. Based on this definition, we conduct a systematic empirical study of state-of-the-art Visual Large Language Models (VLLMs). We find that, under default decoding, VLLMs consistently under-report ambiguity—even when multiple valid visual groundings are present. Importantly, probing model hidden states reveals that ambiguity-related signals are already encoded in their internal representations, despite not being reliably expressed in outputs. Finally, we show that selectively activating multi-focus answering based on these signals can recover additional valid answers while avoiding excessive hallucination. Together, our results suggest that ambiguity in VQA is not merely an annotation artifact or capability limitation, but a property that VLLMs internally recognize yet often fail to surface under standard decoding assumptions.
尚无 AI 解读，展示原文摘要
cs.CL
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表
Qiming Li, Xiaocheng Feng, Yixuan Ma Harbin Institute of Technology, Ruihan Chen, Zihe Tong Harbin Institute of Technology, Zekai Ye Harbin Institute of Technology, Xiachong Feng University of Hong Kong, Libo Qin Central South University 等 13 位作者
Large Language Models (LLMs) and Large Vision-Language Models (LVLMs) demonstrate strong reasoning capabilities, yet their performance in English significantly outperforms that in low-resource languages, raising fairness concerns in multilingual applications. Existing approaches either rely on costly multilingual training or employ prompting with external translation tools, both of which are resource-intensive and sensitive to translation quality. To address these limitations, we propose a training-free inference-time method to enhance Multilingual Reasoning capabilities via Representation Engineering (MRRE) without using any additional training data or tools. MRRE sequentially injects two precomputed vectors at specific layers during inference processing: cross-lingual reasoning enhancement vectors, which steer non-English reasoning representations toward English space to unlock multilingual reasoning, and target-language output anchoring vectors, which restore the distribution of the target language to preserve input–output language consistency. Comprehensive experiments across six advanced LLMs and LVLMs on four reasoning benchmarks demonstrate that MRRE consistently enhances non-English reasoning by an average gain of 5.48% and up to 7.54% in low-resource languages (e.g., Thai and Swahili), while improving input-output language consistency by 3.78%.
尚无 AI 解读，展示原文摘要
cs.CL · cs.CV
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表
Yihan Chen University of the Chinese Academy of Sciences, Jiawei Chen Institute of Software, Chinese Academy of Sciences, Guozhao Mo, Xuanang Chen, Ben He, Xianpei Han Institute of Software, CAS, Le Sun Institute of Software, Chinese Academy of Sciences
The growing use of large language models (LLMs) in peer review threatens scholarly integrity. Recent conference policies allow AI tools for language polishing but prohibit their use for generating substantive content. However, existing detectors mainly rely on stylistic cues, making it difficult to distinguish between surface-level language refinement and genuine content generation. To address this, we advocate a content-based detection paradigm and introduce CoCoNUTS, a comprehensive benchmark containing 315,535 reviews covering leading AI conferences and six human-AI collaboration modes. Our evaluation shows that current detectors struggle to handle these nuanced settings. Consequently, we propose CoCoDet, an AI review detector designed to identify substantive AI-generation. Experiments demonstrate that CoCoDet achieves a macro F1-score of 98.24%. Crucially, on permissible machine-polished reviews, it maintains a low false positive rate of 3.89%, substantially outperforming the strongest baseline (7.84%). Examination on real-world reviews using CoCoDet reveals an escalating trend of substantive AI generation. Our work exposes the inadequacy of current detectors, underscoring the importance of domain-specific solutions.
尚无 AI 解读，展示原文摘要
cs.CL
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表
Ruihan Chen, Qiming Li, Xiaocheng Feng, Weihong Zhong Harbin Institute of Technology, Xiaoliang Yang Harbin Institute of Technology, Yuxuan Gu Harbin Institute of Technology, Zekun Zhou Harbin Institute of Technology, Yunfei Lu Huawei Technologies Ltd. 等 12 位作者
Large Vision–Language Models (LVLMs) have shown strong potential as multilingual Graphical User Interface (GUI) agents, as evidenced by existing GUI benchmarks. However, these benchmarks exhibit two primary limitations: (1) although Perception and Reasoning (P R) capabilities are fundamental for GUI agents, current benchmarks lack fine-grained diagnostics to identify which specific capabilities lead to task failures, hindering targeted improvements; (2) existing benchmarks fail to provide a strictly aligned cross-lingual evaluation environment, introducing confounding factors that prevent isolating the language impact on GUI agent performance. To address these issues, we propose the Multilingual P R GUI Benchmark (MPR-GUI-Bench), featuring strictly aligned environments across six languages and eight fine-grained P R tasks. Our benchmark reveals consistent P R gaps between English and non-English settings, particularly on reasoning-intensive tasks. To leverage the superior English P R capabilities for bridging cross-lingual gaps, we identify layers sensitive to language and propose GUI-XLI, a GUI Cross-Lingual Intervention method that aligns non-English hidden states with their English counterparts at these layers during inference. Experiments show that GUI-XLI effectively reduces the cross-lingual gaps, with an average gain of 6.5% in non-English settings.
尚无 AI 解读，展示原文摘要
cs.CL · cs.CV
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表