Video generation is shifting from end-to-end synthesis toward agentic generation, where planning, tool use, and feedback enable coherent, controllable multi-shot videos. In parallel, video world models provide action-conditioned observations and predicted futures, serving as interactive environments for agents. We study their intersection through the agent harness: an executable runtime that constructs context, exposes tools, maintains state, routes feedback, and manages budgets around video models. This report examines three directions. First, agentic video generation, where harnesses coordinate scripting, shot planning, asset management, production, and editing, alongside the evaluation challenges of these systems. Second, agentic world modeling, where harnesses mediate observation, control, execution, and state. We discuss hybrid designs in which game engines or agent-written programs maintain world state while video models render observations, and how agents and world models can mutually evaluate and improve one another. Third, recursive self-improvement (RSI) in video model development, distinguishing output revisions, updates to harness state, generator parameters, and evaluators, and changes to the improvement procedure itself. We identify how these improvements feed back into subsequent development. Across these directions, we analyze essential agent capabilities, including audiovisual perception, long-horizon planning, memory, code-based interaction, feedback-guided control, skill reuse, and multi-agent coordination. We conclude with challenges in autonomous video production, lifelong memory, programmable hybrid control, autonomous model development, provenance, cost accounting, and evaluation linking video quality and task performance to sustained improvements in models and harnesses.
Yan Tai, Yunan Ding, Guangzhao Li, Siqi Luo, Chaoyang Zhang, Guangtao Zhai, Guodong Guo, Xiaohong Liu
KV cache quantization is well studied for autoregressive language models, where cached prefix states are exact under causal attention and quantization is mainly a cache reconstruction problem. Diffusion language models (dLLMs) break this assumption. Fast dLLM inference reuses KV states across denoising steps, but bidirectional token representations can change as masked tokens are revealed, so the reused cache may already differ from full recomputation before quantization is applied. We study KV cache quantization under this cache-reuse error. We adapt KIVI and KVQuant to Fast-dLLM-style prefix and dual cache reuse, and introduce TRACE and dLLM-KVQ. TRACE reconstructs feasible dLLM reveal trajectories from prompt-answer pairs, providing calibration states that match iterative masked decoding. dLLM-KVQ uses these trajectories to choose per-layer K-cache entries that remain in high precision according to output- and attention-level behavior metrics, while applying standard low-bit value quantization. Across LLaDA and Dream evaluations on GSM8K, HumanEval, MATH, and MBPP, dLLM-KVQ achieves the highest average score among quantized methods across the reported 4-bit and 2-bit prefix/dual cache settings.
World action models (WAMs) increasingly couple internal visual or world representations with action generation, making their world-to-action pathways a critical yet poorly understood source of robustness. We investigate when perturbations to these internal representations actually become behaviorally consequential, and uncover a pronounced mismatch between representation perturbability and closed-loop vulnerability: attack effectiveness varies sharply across visual K/V routes, while larger representation displacement does not reliably imply stronger control failure. Our analysis attributes this mismatch to the action-conditioned readout, where keys determine where action queries read and values determine what content they receive. This motivates an action-facing view of vulnerability, in which the effect of a perturbation depends on how it is routed into action generation rather than on representation-space magnitude alone. We instantiate this view with Tail-AV, a short-probe signal for identifying vulnerable routes, and operationalize it in PASA-KV, which selects action-facing K/V targets before training a fresh task-shared attack. Experiments across multiple WAM systems and a physical robot validate readout-informed target selection, demonstrate strong closed-loop attacks, and reveal substantial system-level differences in resistance.
Ruize Xu, Chenxiao Yang, Yanhong Li, David McAllester
We demonstrate that a small transformer can learn to execute programs in MicroPy, a simplified yet computationally universal programming language. Given procedure definitions together with an expression to evaluate, the transformer predicts small-step execution using PENCIL scaffolding for space-efficient execution within a bounded context window. After training on randomly generated, semantically arbitrary MicroPy programs, the learned transformer generalizes to held-out human-written programs including bit copying and flipping, binary addition and multiplication, and SAT verification and solving. The trained model exhibits out-of-distribution generalization: it correctly evaluates programs drawn from a substantially different distribution than the one used to generate its training data. Since MicroPy can express any computation, our results provide empirical evidence that a standard transformer can be trained to act as a universal computer when paired with PENCIL scaffolding. Project page: https://www.micropy.org
Renzhi Wu, Li Sheng, Zikun Cui, Junjie Yang, Tai Guo, Keke Zhai, Yan Li, Jerry Fu 等 11 位作者
Long-sequence recommendation can improve along two axes: \emph{evidence scaling} exposes longer histories, while \emph{reasoning scaling} adds layers that interpret that evidence for a candidate. Existing models couple them: sequence encoders propagate every added event through depth, whereas one-shot selection prevents later layers from revisiting omitted evidence. We introduce the Latent Workspace Ranker (LWR), which keeps typed evidence in persistent banks and propagates a fixed-size candidate-local workspace. Each layer rereads the banks from its evolving workspace and archives compact post-interaction state, allowing history length and reasoning depth to grow independently. We compare four architectures across various history--depth configurations using evidence scaling at fixed depth, reasoning scaling at fixed 10K history, and their joint compute-efficient frontier. LWR leads all families at long histories, continues improving with reasoning compute, and converts additional training compute into quality gains at the fastest fitted rate on the joint frontier. At 10K history, it improves NE by 0.19% over HSTU while using 69.6% fewer forward GFLOP/request. It also outperforms existing baselines on two billion-scale ranking surfaces. These results show that separating addressable evidence from propagated state makes reading more and reasoning deeper independently controllable and jointly useful.
Renzhi Wu, Zikun Cui, Junjie Yang, Li Sheng, Yongjun Xie, Jerry Fu, Nipun Mathur, Sri Reddy 等 9 位作者
Sequential recommenders typically construct history inputs from item identifiers or content features using an embedding map shared across users. The same item therefore yields the same initial embedding representation, even when users have different relationships to the item. We propose personalized item representations in user history, where each item's embedding expresses its relationship to the user. Personalization becomes a property of the history inputs themselves, supplying richer evidence for later sequence modeling. We instantiate this idea with PRISM, which reuses upstream ranking computation to generate personalized item embeddings. An autoencoder jointly compresses all layer states from a user tower, a user--item tower, and same-request candidate mixing into one embedding per scored pair for downstream replay. Offline evaluation on Surface 1, a major recommendation surface at billion scale, reports an average downstream normalized entropy (NE) change of $-0.15\%$, compared with $-0.07\%$ for VISTA at similar history lengths. We further identify an empirical sequence-length scaling law: over history lengths of 26--240 events, NE gain follows a fitted power law with exponent 0.468, increasing from 0.06\% to 0.17\%. Therefore, personalized item representations in user history provide a new scaling dimension for improving recommendation quality.
Hong Lin, Zhongle Xie, Lidan Shou, Ke Chen, Gang Chen
Customizing Pretrained Models (PMs) for downstream tasks demands task-specific data, which burgeoning Web3 applications provide abundantly. Blockchain-powered federated learning (BCFL) offers a secure and privacy-preserving paradigm to customize PMs among self-interested and untrusted Web3 participants. Yet, existing BCFL algorithms face severe efficiency and scalability challenges when applied to large PMs, due to their substantial computational and memory requirements and the hardware constraints of many participants. This paper introduces DICE, a novel BCFL algorithm that addresses the efficiency and scalability challenges of customizing large PMs in Web3 via two sets of synergistic techniques: (i) a dynamic progressive learning scheme and a model slice verification scheme to enable a wider range of honest participants, including those unable to deploy the full model, to engage in customizing PMs, while preserving model utility; (ii) a tree-structured committee consensus protocol and an efficiency-driven incentive scheme to address the dominant bottleneck that hinders existing BCFL algorithms from scaling to a larger number of data holders for customizing data-hungry large PMs, while maintaining security. Theoretical analysis and extensive experimental results on customizing vision and language PMs demonstrate DICE's efficiency, effectiveness, and security.
Zhefan Rao, Meng Chu, Xuan Billy Zhang, Jingfan Chen, Guohao Cai, Runyi Li, Shiyi Du, Xu Huang 等 92 位作者
World Models (WMs) can support Multimodal Agents (MAs) through simulated experience, while Agent trajectories can inform World Model refinement. Evaluating this interaction requires separate evidence about action selection and state evolution. We introduce Dual Exam, which tests Agents under fixed environment response mechanisms and World Models under fixed action interventions across physical, digital, social, and scientific domains. The Agent track compares 10 configurations on 1,088 items in 15 reporting categories. The World Model track contains 400 cases across eight scenarios; 15 configurations have full coverage and one covers six scenarios. GPT 6 Astra leads the case-weighted Agent Index at 31.3, and GPT-Image-2 leads the full-coverage WM Overall at 21.7. These scores use different tasks and metrics and are not directly comparable. A 150-case pilot of one coupled pairing records zero task successes and motivates component-level diagnosis. Results reveal uneven task performance and failures to preserve task-relevant state. Their value for predicting coupled success or sustained reciprocal learning remains to be established.
Vision-language models are shifting from offline passive understanding toward online proactive interaction. However, prevailing methods largely rely on explicit requests while overlooking implicit user intents (e.g., "where should I click next?" when a user is navigating a website) and fail to continuously update their understanding as user intents evolve. These limitations make them less suitable for intent-aware interaction in real-world scenarios. To address these challenges, we introduce StreamVibe-Bench, a new benchmark that serves as a testbed for user intent understanding and interaction, comprising 4.1K evaluation samples and 85K training samples spanning explicit, implicit, and dynamic intents. Building on this benchmark, we develop StreamVibe-VL, a streaming proactive model that infers implicit user intents from continuous observations and dynamically adapts its interactions as these intents evolve. We further introduce StreamVibe-Harness, which maintains user intents as persistent states with flexible lifecycles through hierarchical memory, enabling it to track the latest user intents and environmental states. Our experiments reveal three key findings: (i) StreamVibe-VL preserves strong general video understanding capabilities while improving response timing and answer quality by 10.7% over existing baselines; (ii) existing proactive models remain limited in understanding user intents and determining when to respond, especially under implicit and dynamic intents; and (iii) StreamVibe-Harness further improves response consistency and adaptation during sustained interactions, resulting in a 12.3% gain in proactive interaction performance.
Zhuoming Liu, Jinhong Lin, Kwan Man Cheng, Lin Zhang, Shayok Bagchi, Yin Li
Many modern vision-language models (VLMs) build on autoregressive decoding of discrete tokens. While text-based output interfaces enable scalable pretraining and strong zero-shot generalization across diverse tasks, they are poorly suited for problems that require precise continuous outputs, such as localizing temporal boundaries of events or generating robotic control actions. To address this challenge, we propose DRIFT, a general framework for adapting pretrained VLMs to continuous decoding tasks. DRIFT combines a base predictor, which provides a coarse estimate of the target output, with a generative refinement module based on flow matching that iteratively improves the prediction. This residual formulation transforms the generative modeling problem from learning a global output distribution to modeling a localized residual distribution around a strong prior, substantially simplifying optimization. We evaluate DRIFT on both perception and planning tasks, including visual grounding and robotic control. Across multiple tasks and architectures spanning MLLMs, VLAs, and WAMs, DRIFT consistently outperforms a strong set of regression- and generative-based solutions.
In reciprocity-based over-the-air split learning, channel aging rotates and attenuates the derivative signal returned to the transmitter. Seven case studies combine numerical forecasts, mechanism checks, and competing predictions. Near the correlation zero, an archived SGD comparison finds truncated inversion 0.000813 lower in cross-entropy than polar correction, with a paired 95% interval for the difference from -0.001853 to 0.000060, and a retained-mode quadratic proxy slightly favors truncation. Selecting the return port with the largest received power falls short of the predicted five-point alignment advantage over random selection at both tested spacings, gaining at most 0.97 points, while the local inverse relation between amplitude and phase gradient is supported. Seven specified interference-induced accuracy-collapse predictions are rejected, and independent zero-mean interference preserves the conditional mean under a fixed linear chain. Per-subcarrier correction and spherical-wavefront compensation do not deliver their predicted accuracy gains. Focusing leaves both methods near chance in a single-mode model and adds no observed gain in a model with a line-of-sight component plus scattering. A successful port assignment under nonsymmetric motion refutes an exact-symmetry requirement. The competing positioning-error prediction is supported. Gaussian rail jitter with a standard deviation of 0.383 wavelengths yields high terminal test accuracy, whereas a fixed bias of the same magnitude yields near-chance accuracy. These cases distinguish valid local physical relations from unsupported training forecasts through task support, noise injection, and update moments.
A Gaussian random wave field is sampled on a lattice within a bounded aperture. After displacement, re-indexing leaves a residual in the Voronoi cell. We study whether the empirical lag correlation has positive real part throughout the cell. For each realization in a single-output reciprocal link with noiseless over-the-air return, this is equivalent to the negative returned gradient being a descent direction for all residuals and differentiable losses with nonzero error signals. Return noise is treated exactly. For isotropic fields, ball-shaped apertures, unshifted lattice points, and dilations approaching the critical covering radius, every minimizer is a deep hole except with exponentially small probability. A Gaussian vector on the deep holes determines the limiting failure probability, with explicit covariance and lattice-independent limiting normalization. Square, hexagonal, simple cubic, and body-centered cubic lattices have two, three, four, and twelve sign classes. At fixed target failure probabilities below the critical limiting value, this covariance determines the exact leading aperture constant. When the margin and target failure probability vanish jointly, that constant is common to lattice shapes within each dimension. A cubic correction yields vanishing relative tail error in a moderate-deviation regime. A non-asymptotic two-sided aperture law has explicit constants. With a Rayleigh-faded specular path and positive diffuse margin, a failure floor occurs at every positive path power exactly when the cosine of its phase change is negative somewhere in the cell. The floor vanishes for all directions exactly when the covering radius is at most a quarter wavelength.
Diffusion models have achieved remarkable success in text-to-image generation, yet aligning their outputs with human preferences remains challenging. Diffusion-DPO learns from image preference pairs, but their global labels do not specify how supervision should be allocated across spatial locations and denoising timesteps. The semantic roles of prompt tokens are also not explicitly considered when assigning these local weights. To address these limitations, we propose Token-Oriented Preference Optimization (ToPO), which builds a token-conditioned spatial-temporal weighting map for each pair. A frozen reference denoiser obtains spatial and temporal weights from local noise-prediction error differences under shared noise, while preferred-branch cross-attention links spatial evidence to content tokens. The map reweights Diffusion-DPO and supports an auxiliary pixel-midpoint ordering loss. It requires neither local preference annotations nor a separately trained reward model. Extensive experiments across multiple preference and compositional evaluations show that ToPO achieves state-of-the-art performance among the compared methods with only 500 training updates, demonstrating the potential of token-conditioned preference allocation for high-quality diffusion alignment within a limited update budget.
Runsheng Chen, Yihong Liu, Raoyuan Zhao, Michael A. Hedderich, Hinrich Schütze, Matthias Grabmair
On-policy distillation can transfer the reasoning capabilities of a strong teacher to a smaller student. However, distillation from a stronger teacher model faces two mismatches: its solutions may be difficult for a smaller student to imitate, and its token-level probabilities may be incompatible with, or even unavailable to, the student. Existing methods handle these settings through different forms of supervision. To this end, we propose Bridged Self-Distillation (BSD), a unified framework for distilling from any teacher that provides reasoning traces. The model to be distilled from, termed the external teacher, first provides a reasoning trace as privileged context, which can be generated in an offline fashion. The student generates rollouts without this context, while a copy of the student conditioned on it, serving as the self-teacher, scores the same rollouts in the student's own token space. BSD uses the resulting token-level guidance to adjust updates on unsuccessful rollouts, with a verifier providing the outcome signal. As the external teacher is queried only to produce text, the same training procedure applies to all teacher scenarios: shared-tokenizer, cross-tokenizer, and closed API teachers. Across these scenarios, BSD improves over the base model and outperforms baselines tailored to each scenario, offering a novel and unified pathway for distillation. Code is available: https://github.com/TUMLegalTech/Bridged-Self-Distillation
Bimanual robot manipulation plays an important role in real-world robotic applications, but the high cost of collecting bimanual demonstrations limits the scalable development of bimanual policy learning. Meanwhile, existing single-arm manipulation data are difficult to directly transfer to bimanual systems due to the Single-to-Dual gap in action spaces, coordination requirements, and arm-side spatial preference. To address this issue, we propose RoboSTD, a zero-shot Single-to-Dual data construction method for bimanual robot learning. RoboSTD first applies sagittal-plane mirroring to observations, proprioception, and actions to construct executable contralateral pseudo-demonstrations. It then uses an LLM to generate structured spatio-temporal coordination constraints that guide the reconstruction of pseudo-bimanual supervision across workspace regions and task stages. We evaluate RoboSTD on the AgileX ALOHA platform in both RoboTwin simulation and real-world settings. Experiments demonstrate the effectiveness of RoboSTD for bimanual policy learning. Without requiring additional bimanual demonstrations, RoboSTD increases the average number of collected cups to 2.11x that of the single-arm baseline in the cup collection task. Project page: https://huahungy.github.io/RoboSTD/
Individual-level genotype data underpin genetic discovery, disease-risk modeling, and therapeutic development, yet many critical settings remain data-limited, including underrepresented ancestries, rare or imbalanced disease cohorts, and restricted-access biobanks. Conditional synthetic genome generation would support privacy-aware benchmarking, simulation, and data augmentation, but requires learning an ultra-high-dimensional discrete distribution over hundreds of thousands to tens of millions of correlated single-nucleotide polymorphisms (SNPs), or sites of DNA variation. We introduce HELICS: High-fidelity gEneration via Latent Interpolant Conditional Synthesis, a phasing-free latent flow matching framework for genome-wide SNP-level generation from biobank-scale genotype data. HELICS tokenizes chromosomes with convolutional autoencoders, reducing >600K SNPs to 4,096 continuous latent tokens, and trains a transformer parameterized conditional flow matching model over the concatenated multi-chromosome latent representation. Trained on over 417K genomes in the UK Biobank, HELICS generates genome-wide common-variant synthetic cohorts without haplotype phasing and achieves stronger fidelity to real genotypes than existing haplotype-based simulators, including high allele-frequency agreement in external evaluation and superior preservation of local and long-range SNP correlation structure. Conditioning on ancestry and genetic liability summaries enables targeted generation across genetic risk profiles; in perturbation analyses, by increasing the conditioned disease trait liability, HELICS induces SNP-level dosage changes in generated genomes that are highly correlated with published association weights. HELICS provides a scalable generative modeling framework for SNP-level genome data and a path toward conditional synthetic cohorts for data-limited genetic studies.
Extending large language models (LLMs) to the speech domain has recently gained significant attention. A typical approach connects a pretrained LLM with an audio encoder through a projection module and trains the resulting model on large-scale, task-specific instruction-tuning datasets. However, curating such instruction-tuning data for specific requirements is time-consuming, and models trained in this manner often generalize poorly to unseen tasks. In this work, we first formulate that the strongest generalization of a speech-LLM is achieved when it is trained with Self-Generated Instruction-Free Tuning (SIFT), in which supervision signals are generated by a frozen LLM using textual representations of speech as input. Our proposed SIFT paradigm eliminates the need for collecting task-specific question-answer pairs and yields the theoretically best generalization to unseen tasks. Building upon this paradigm, we introduce AZeroS (Auden Zero-instruction-tuned Speech-LLM), which is trained on speech-text pairs derived from publicly available corpora, including approximately 25,000 hours of speech with ASR transcripts and 3,000 hours of speech with paralinguistic labels. Built upon Qwen2.5-7B-Instruct, the model updates only two lightweight projection modules (23.8 million parameters each), while keeping both the LLM and audio encoders frozen. Despite the minimal training cost and modest data scale, AZeroS achieves state-of-the-art performance on both semantic and paralinguistic benchmarks, including VoiceBench, AIR-Bench Foundation (Speech), and AIR-Bench Chat (Speech).
Yijun Zhou, Minghao Shao, Ke Zhang, Renkai Yuan, Hongbo Yuan, Bin Hu, Bozhao Qian, Hongxia Shen 等 23 位作者
Game development stresses coding agents in ways general software engineering does not: correctness depends on how engine APIs, serialized scene state, and runtime behavior interact, not on source code alone. Existing benchmarks mainly target platforms such as Godot and heavily rely on LLM judges or coarse task-level outcomes. We present UnityBench-Mini3D, a benchmark of 77 project-level Unity tasks organized around seven expert-defined skill categories rather than tutorials or genres, with three difficulty tiers. Each task pairs a complete Unity project with natural-language requirements and is evaluated using 1,178 held-out, human-authored milestone tests covering scene structure, component configuration, and runtime behavior. This in-engine evaluation provides deterministic partial credit, localizes failures, and distinguishes partial implementations from complete solutions without relying on subjective judges. Across 21 models and 9 coding harnesses, the average pass rate reaches 80.9% on easy tasks but falls to 35.2% on hard tasks, with just 57.7% overall. UnityBench-Mini3D provides a reproducible, diagnostic testbed distinguishing systems across difficulty levels.
In over-the-air split learning over an aging single-antenna Rayleigh channel, the mean returned gradient equals the antenna-domain gradient scaled by the temporal correlation. Uncorrected training collapses in negative-correlation lobes. Near a correlation zero, no nonzero scalar calibration improves the ratio of mean power to innovation power for an individual transmission, and, on the tested grid, using sixteen times as many pilots recovers only 7 percent of the training gap. Returning at the first correlation extremum with phase correction, 1.3 ms later at 3.5 GHz and 15 m/s, achieves 90.2 percent accuracy, compared with 91.8 percent with CSI. In the tested line-of-sight setting, the near-zero gap becomes negligible.
Over-the-air split learning returns the receiver's loss gradient through the reciprocal channel, so terminal motion between the forward and backward transmissions corrupts the backward pass. We repair the aged return at a fixed receive array using only second-order statistics. Two numbers taken from the forward-backward cross-covariance, the alignment reachable by the best common phase and by the best unitary correction, bound the main term of every repair. In the 3GPP tapped-delay-line channels TDL-A/B/C, where every antenna ages by the same scalar, the two numbers coincide, and at a correlation zero all tested fixed corrections stay between 58% and 62% accuracy, whereas a return delayed to the next correlation lobe reaches 98.4%. In 3GPP clustered-delay-line channels a statistical unitary correction stays within 0.02 in mean cross-entropy of ideal channel knowledge in all six tested profile and speed combinations, while common-phase correction falls to 69% for CDL-A at 3 km/h and the tested delay loses up to 10 points. In a two-cluster model spatial correction is decisive only where the common-phase alignment nearly vanishes (64% to 95%). With a ResNet-18 on CIFAR-10, only the delayed return trains at a separable correlation zero (86% against at most 16%), and at 30 km/h in a clustered-delay-line channel the unitary correction reaches 80% where the common phase reaches 60%. Given geometry, displacement and angular statistics, a port permutation chosen by linear assignment reaches 99.5-99.7% accuracy for hexagonal and square arrays in oblique motion without pilots or digital feedback during training, where nearest-position matching drops to 89%.
Yuhang Dong, Min Xia, Jinghong Li, Yue Dai, Xinyue Wang, Shi Liang
Wind power forecasting models are often deployed as fixed temporal backbones, but the errors they make can vary across meteorological regimes such as cold waves, heat-and-calm periods, productive typhoon, and destructive storms. We present a guarded Agent-in-the-loop residual adaptation framework that leaves the temporal backbone unchanged and adapts only a small set of scene-conditioned residual experts. An Agent layer interprets external scene information, identifies the current regime, selects residual experts, and applies a correction only when recent observed errors support it. New experts enter the pool after they outperform their initializations on a validation set, keeping model growth under explicit control. Backbone-only and Agent-enhanced forecasts are compared at lead times of 4, 8, 12, 24, and 48 across nine time-series backbones. Across these runs, the guarded framework lowers both MAE and RMSE for all 45 backbone-horizon combinations, with average MAE reductions ranging from 11.21% to 20.26% across backbones. Online-selection baselines, component ablations, and routing analyses identify the roles of residual experts, scene-aware routing, and the online acceptance rule. Candidate-scene runs evaluate the promotion path under controlled conditions; long-term unattended deployment was not evaluated.
Yuhang Dong, Jinghong Li, Xinyue Wang, Yuanning Cui
Fact-verification benchmarks score the final label, while an interactive verifier must first decide whether the available evidence is sufficient. We focus on that intermediate Stop-or-Clarify Decision. Uncertainty alone is insufficient, because the benefit of another evidence-acquisition step is unknown before the step is taken. Sequential Experience-Adaptive Clarification (SEAC) wraps a fixed verifier with a lightweight evidence-acquisition policy. For each claim, SEAC retrieves evidence, runs the verifier, forms a label-free decision state, and combines stop risk with an observational clarification-utility estimate from earlier executed trajectories. Clarification is requested only when this score exceeds the action cost and decision threshold. On SciFact, LIAR, and FEVER, SEAC improves mean accuracy and macro-F1 over the compared fact-verification and active-retrieval baselines. Equal-budget and formula-level diagnostics support selective clarification as a contributor to the gain, while the small SciFact split leaves its paired improvement inconclusive.
Boyuan Meng, Peihua Bao, Hong Liu, Xiaowei Zhu, Chao Wang, Gen Li, Zhenxuan Pan
Agentic reinforcement learning (RL) generates irregular rollout trees, but trajectory-by-trajectory training recomputes their shared interaction histories. Existing training-time prefix-reuse systems do not jointly address state recovery, dense execution, and distributed training for hybrid-attention models. We present HARTS (Hybrid-Attention RL over Tree Structures), to our knowledge the first system to demonstrate training acceleration from arbitrary rollout-tree prefix sharing on real hybrid-attention models. A two-level planner constructs compression-oriented microbatches, then jointly refines their membership, cross-microbatch sharing, placement, and execution order. Within each sharing window, gradients from dependent microbatches accumulate before one backward pass through the shared prefix computation. Differentiable boundary-state handoff and bounded replay preserve the per-trajectory chunkwise numerical path, while a linear-time planner attains the global minimum number of sequential linear-attention calls under our packed chunkwise-call model. The runtime uses a common compact layout across attention and dense or mixture-of-experts (MoE) layers, with support for distributed parallelism and activation recomputation. Recovering log-probabilities for each original trajectory position preserves RL objectives; for deterministic top-k routing without token dropping, occurrence counts restore MoE objective weights and expert-load statistics. On SWE-bench, HARTS delivers 4.77-10.45x the update throughput of trajectory-expanded training across two sampling schemes on 7.9B- and 124B-parameter hybrid-attention MoE models using 8-64 GPUs. On the 7.9B model, full-model logit differences are comparable to baseline self-rerun variation. The 124B model shows stable online learning trends through step 150 on SWE-bench and $τ^3$-Bench.
Yuheng Zha, Yilei Wang, Qiyue Gao, Junrong Chen, Yujia Wu, Zhengfeng Lai, Zhengzhong Liu, Eric P. Xing
Humans often solve spatial problems by mentally simulating visual transformations. In contrast, conventional vision-language models (VLMs) reason primarily through language. We investigate whether VLMs can solve spatial problems by reasoning with both text and generated visual states. To this end, we introduce WM-VLM, which equips a pretrained VLM with a lightweight world model branch for generating intermediate visual states. Our two-stage training first teaches the model to generate the next visual state and then to use that state for reasoning. We programmatically construct spatial reasoning tasks with verifiable intermediate visual states. These tasks allow us to evaluate how well the model generates visual states and how much it relies on them to answer the question. On 2D and 3D mental rotation tasks, WM-VLM consistently outperforms the supervised fine-tuned backbone, with gains of up to 39.25 percentage points. Ablations suggest that these gains depend on the generated visual states, as removing or corrupting them sharply reduces performance. Together, these results suggest that internal world models offer a promising path toward VLMs that reason in both language and visual space. Code is publicly available at: https://github.com/yuh-zha/WM-VLM
Yingzhuo Deng, Chong Li, Bingxin Liu, Yi Shu, Tianyu Peng, Can Xie, Siyao Song, Shuo Ren 等 9 位作者
Spoken Dialogue Models (SDMs) typically share a speech-token space for generation and synthesis, leading to long autoregressive sequences and costly inference. Lower-rate speech tokenizers shorten sequences but may introduce compression-quality trade-offs and costly replacement without breaking this coupling. To address this limitation, we propose CoSTA, a compatibility-preserving post-hoc framework that decouples generation and synthesis token spaces while retaining the original speech tokenizer and decoder. CoSTA extends the generation token space with compressed representations via mean-initialized vocabulary extension, pronunciation adaptation, and dialogue adaptation, followed by deterministic expansion back to the original speech-token space before synthesis. On bilingual speech-dialogue benchmarks, CoSTA achieves approximately $1.7\times$ expanded speech-token generation speed, lower full-response latency, and higher throughput while preserving response quality. Analyses show more sustained interaction under limited budgets and a practical trade-off among compression, decoding efficiency, and quality. Further experiments demonstrate cross-architecture transfer and complementarity to low-rate tokenization.
Kejuan Yang, Yizhuo Zhang, Mingyuan Du, Yue Zhang, Dixin Zheng, Kaili Zhao, Yang Xiao, Hanzhong Liang 等 9 位作者
Global-scale video moderation faces a dual challenge: the need for fine-grained multimodal reasoning and the demand for interpretable outputs to support downstream enforcement. Traditional moderation systems often rely on fragmented black-box classifiers that are difficult to maintain and lack transparency.In this paper, we present UNIVID, a Unified Vision-Language model for Video Moderation. Unlike standard classification models, UNIVID generates policy-aware captions that serve as an interpretable intermediate representation, enabling human-verifiable decisions and multi-task reusability. While existing open-source and commercial VLMs often suffer from safety-guardrail refusals and lack fine-grained policy alignment, we develop a specialized training data recipe that combines expert human-refined labels with synthetic data to align the model with our safety guidelines.By integrating UNIVID as the core captioner, we design a novel end-to-end video moderation system that reduces violation leakage by 42.7% and overkill rate by 37.0% relatively. Meanwhile, by replacing over 1,000 policy-specific models with a single UNIVID backbone, we recycle extensive computational resources while significantly reducing engineering maintenance overhead. To our knowledge, this is one of the first reports of a high-efficiency captioning VLM successfully supporting industrial-scale moderation and cross-functional business.
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 6: Industry Track) · 2026 · 正式发表
Mingzhe Lu, Yiwen Wang, Yanbing Liu Institute of Information Engineering，Chinese Academy of Sciences, Qi You, Chong Liu University of Science and Technology of China, Ruize Qin, Haoyu Dong, Wenyu Zhang Baidu 等 11 位作者
Computational narrative analysis aims to capture rhythm, tension, and emotional dynamics in literary texts. Existing large language models can generate long stories but overly focus on causal coherence, neglecting the complex story arcs and orchestration inherent in human narratives. This suggests a structural misalignment between model- and human-generated narratives.We therefore position narrative analysis as a diagnostic proxy for generation and propose VISTA Space, a high-dimensional framework for narrative orchestration that unifies human and model perspectives while jointly characterizing narrative function and structure in a common space.We further introduce LitVISTA, a structurally annotated benchmark grounded in literary texts, which operationalizes VISTA Space for systematic evaluation of models’ narrative orchestration capabilities. Under an oracle setting with gold event anchors, we evaluate frontier LLMs including GPT, Claude, Grok, and Gemini. Results reveal systematic deficiencies, as current models struggle to jointly capture narrative function and structure and fail to form an integrated global view of literary narrative orchestration. End-to-end analysis further shows that failures are dominated by anchor identification and localization errors. Even advanced thinking modes yield mixed and often limited gains for literary narrative understanding.
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
Qianli Wang, Van Bach Nguyen Phillips-Universität Marburg, Yihong Liu Ludwig-Maximilians-Universität München, Fedor Splitt Technische Universität Berlin, Nils Feldhus, Christin Seifert Phillips-Universität Marburg, Hinrich Schuetze, Sebastian Möller 等 9 位作者
Counterfactuals refer to minimally edited inputs that cause a model’s prediction to change, serving as a promising approach to explaining the model’s behavior. Large language models (LLMs) excel at generating English counterfactuals and demonstrate multilingual proficiency. However, their effectiveness in generating multilingual counterfactuals remains unclear. To this end, we conduct a comprehensive study on multilingual counterfactuals. We first conduct automatic evaluations on both directly generated counterfactuals in the target languages and those derived via English translation across six languages. Although translation-based counterfactuals offer higher validity than their directly generated counterparts, they demand substantially more modifications and still fall short of matching the quality of the original English counterfactuals. Second, we find the patterns of edits applied to high-resource European-language counterfactuals to be remarkably similar, suggesting that cross-lingual perturbations follow common strategic principles. Third, we reveal that multilingual counterfactual data augmentation (CDA) yields larger model performance improvements than cross-lingual CDA, especially for lower-resource languages. Yet, the imperfections of the generated counterfactuals limit gains in model performance and robustness. Finally, we identify and categorize four main types of errors that consistently appear in the generated counterfactuals across languages.
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
As the post-training of large language models (LLMs) advances from instruction-following to complex reasoning tasks, understanding how different data affect finetuning dynamics remains largely unexplored. In this paper, we present a spectral analysis of layer-wise gradients induced by low/high-quality instruction and reasoning data for LLM post-training. Our analysis reveals that widely-studied metrics for data evaluation, e.g., IFD, InsTag, Difficulty, and Reward, can be explained and unified by spectral properties computed from gradients’ singular value decomposition (SVD). Specifically, higher-quality data are usually associated with lower nuclear norms and higher effective ranks. Notably, effective rank exhibits better robustness and resolution than nuclear norm in capturing subtle quality differences. For example, reasoning data achieves substantially higher effective ranks than instruction data, implying richer gradient structures on more complex tasks. Our experiments also highlight that models within the same family share similar gradient patterns regardless of their sizes, whereas different model families diverge significantly. Providing a unified view on the effects of data quality across instruction and reasoning data, this work illuminates the interplay between data quality and training stability, shedding novel insights into developing better data exploration strategies for post-training.
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
Ming Li, Pei Chen Amazon, Zhenhao Zhang Amazon, Tao Yang Amazon, Xinyang Zhang Anthropic, Han Li Shopify, Tianyu Cao, Ming Zeng Amazon 等 13 位作者
Large Language Models demonstrate strong capabilities in single-turn instruction following but suffer from Lost-in-Conversation (LiC), a degradation in performance as information is revealed progressively in multi-turn settings. Motivated by the current progress on Reinforcement Learning with Verifiable Rewards (RLVR), we propose Curriculum Reinforcement Learning with Verifiable Accuracy and Abstention Rewards (RLAAR), a framework that encourages models not only to generate correct answers, but also to judge the solvability of questions in the multi-turn conversation setting. Our approach employs a competence-gated curriculum that incrementally increases dialogue difficulty (in terms of instruction shards), stabilizing training while promoting reliability. Using multi-turn, on-policy rollouts and a mixed-reward system, RLAAR teaches models to balance problem-solving with informed abstention, reducing premature answering behaviors that cause LiC. Evaluated on LiC benchmarks, RLAAR significantly mitigates LiC performance decay (62.6% to 75.1%) and improves calibrated abstention rates (33.5% to 73.4%). Together, these results provide a practical recipe for building multi-turn reliable and trustworthy LLMs.
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