每天知道
你的领域
出了什么

Inference-time activation steering enables behavioral control of large language models without parameter modification, while post-training quantization reduces memory and compute costs for deployment. Despite their growing convergence in practice, the interaction between these two techniques remains uncharacterized. We systematically study activation steering under weight-only quantization (INT8 and NF4) across four open-weight 7-9B models and two behavioral targets: judged sentiment and judge-free reasoning length. Using an iso-effect framework that compares capability costs at matched behavioral effect, we find that sentiment steering survives quantization intact. After correcting a GSM8K parser artifact with a uniform v2.3.1 rescore, the pooled INT8 contrast is -0.010 (90% CI [-0.026, +0.007]), descriptively Equivalent under the preregistered three-label rule, while NF4 remains Inconclusive at -0.017 ([-0.067, +0.033]). In contrast, reasoning length exhibits a surprising asymmetric dose-response: lengthening is graded but terminates in cap-runaway and collapse, while shortening is a step function with only 12-30% shortening (model-dependent) before discontinuous failure. We expose a methodological pitfall: the naive iso-effect ladder anchors on the collapse floor for floor-bounded targets, and we introduce a censored construction that restores interpretable crossings. We also quantify a substantial baseline capability shift for Mistral-NF4 (0.545 to 0.365 GSM8K at alpha=0), demonstrating that compression can dominate the steering intervention. Despite this, steering vectors remain highly collinear with their FP16 siblings (cosine similarity 0.989-0.998 for INT8, 0.945-0.990 for NF4), confirming that the behavioral direction survives quantization even when the cost structure does not. All code and data are released.

Although multimodal Large Language Models (MLLMs) excel in diverse tasks, their scalability remains limited by the memory and computational overhead of KV cache storage. Recent KV cache eviction approaches incorporate a cosine similarity-based diversity metric with importance metrics to selectively retain critical key-value pairs. However, cosine similarity involves normalization that discards magnitude information, and it often yields uniformly high similarity values across layers due to the anisotropy property of hidden representations. In our study ECOKV, we rigorously deconstruct the capabilities of existing diversity metrics. Moving beyond simple measurement, we propose a geometry-aware composite metric that jointly leverages Euclidean distance and cosine similarity to capture token diversity from complementary perspectives. Furthermore, we use these two metrics to estimate the redundancy level of each attention head, allowing adaptive weighting between diversity and importance scores during token selection. Finally, we demonstrate that the observation window commonly employed to preserve recent tokens can be substantially reduced, thereby allocating more cache capacity to informative tokens and yielding consistent improvements. Extensive experiments demonstrate that ECOKV achieves state-of-the-art performance under various compression ratios and can be seamlessly integrated with existing KV cache eviction methods. We further analyze the relationship between importance and diversity, and examine redundancy patterns across layers and attention heads.

Shared key--value (KV) cache reuse improves large language model (LLM) serving, but it can also create a timing side channel that reveals whether a prefix is already cached. Previous work shows that such attacks are possible, but their reliability under realistic multi-tenant contention is less understood. We study this problem through seven experiments on live shared LLM-serving systems. On a vLLM server running DeepSeek-R1-Distill-Llama-8B on NVIDIA GB10, mean Cohen's d drops from 0.7789 with no synthetic workers to 0.2109 with two workers (t=8.412), while higher worker counts cause no statistically detectable further loss. A 120-run sparse-overlap experiment places the best breakpoint at the boundary of the measured range (tau=0, 95% CI [0.000,0.113]), supporting an ambient-versus-loaded regime change rather than an internal physical threshold. AUROC falls from 0.650 at ambient to 0.531 near 61% overlap and partially recovers to 0.574 at saturation. Concurrency-depth variance is the strongest measured correlate of effect size (r=-0.416) and hit consistency (r=-0.637). An interleaved control preserves the same non-monotonic ordering. The main collapse is also reproduced on a real two-node, two-GPU tensor-parallel vLLM setup, where mean d falls from 3.418 to 0.511 (p<0.01). Two SGLang pilots are statistically inconclusive. Overall, KV-cache timing reliability depends strongly on the load regime and serving stack, and measurements on quiet systems can overestimate operational attack reliability.

With the widespread applications of large language models (LLMs), privacy-preserving inference has become increasingly essential for sensitive queries. To balance privacy and utility, a series of lightweight obfuscation approaches has recently been proposed, where users locally transform plaintext embeddings into the fixed ciphertext ones. While such Embedding-to-Embedding Obfuscation (E2EO) schemes demonstrate considerable resilience against traditional token frequency and embedding inversion attacks, the core mechanism behind remains to be the large-scale one-to-one substitution, which provides no cryptographic guarantees. In this paper, we propose Proxy Manifold Alignment (PMA), a novel attack against E2EO in privacy-preserving LLMs. Our key observation is that E2EO schemes keep the original semantic structure, so that the obfuscated vector stream can be regarded as an unknown tokenizer-language whose symbols are the vectors themselves. Therefore, the proposed ciphertext to plaintext reconstruction attack can be formulated as a translation task from the unknown tokenizer-language to plaintext. Specifically, by only accessing the obfuscated vector stream, the target tokenizer and a public corpus, the PMA attack first employs Word2Vec to model the co-occurrence patterns within the obfuscated stream and the public corpus independently, and constructs two proxy vector embeddings. Then, the attack aligns the underlying manifolds of these two embeddings based on structural similarity. Finally, it maps the obfuscated vectors back to plaintext. Experimental results demonstrate that PMA consistently achieves higher plaintext recovery than other state-of-the-art attack methods.

Cross-lingual alignment (CLA) aims to align the representations of large language models (LLMs) across languages, enabling cross-lingual transfer to improve multilingual capabilities. Previous CLA methods often ignore language-specific information encoded in representations and only consider sentence-level alignment, which may lead to suboptimal performance and input-output language mismatch. We propose CAROT (Cross-Lingual Alignment of Representations in a Language-Agnostic Space via Optimal Transport), which consists of two steps: identifying language-specific representations in LLMs' internal states and aligning language-agnostic representations across languages at the token level by optimal transport, while explicitly preserving language-specific representations. Inference-time steering experiments show that the representations computed by CAROT are effective alignment targets, improving multilingual performance by up to 11.2 points in accuracy while maintaining input-output language consistency. We further use the representations obtained by CAROT as training targets, internalizing the aligned representations. The trained models outperform existing CLA methods in 11 of 18 evaluation settings (3 models $\times$ 3 tasks $\times$ ID/OOD languages). Our work provides insights into what constitutes effective alignment targets for CLA in LLMs. Code is available at https://github.com/ynklab/CAROT

The next generation of artificial intelligence systems will likely be natively temporal and multimodal in both inputs and outputs: able to converse, perceive, predict, reason, and synthesize through a shared world representation. Realizing this requires a temporal perceptual substrate integrating sensory streams, language, and structured outputs within a multimodal world model. We seek a generalist open-world perceptual system that represents biological forms, natural physical structures, and artifacts, and their interactions, as a coherent, temporally persistent process. The model should infer geometry, articulation, semantics, interaction structure, and uncertainty from raw multimodal streams; maintain identity through occlusion and viewpoint change; generalize across species, forms, mechanisms, and materials; and abstain or expand its ontology when encountering the unknown. The objective is a structured world state supporting understanding, prediction, counterfactual reasoning, and controllable synthesis. Recent work suggests that some cross-modal and reasoning-like capabilities can emerge from large-scale generative video pretraining, reminiscent of language-model scaling. Yet these capabilities are often accessed through language probes or expressed through photorealistic video, leaving explicit semantic, geometric, or temporal structure largely unexposed. This paper articulates an alternative and complementary paradigm: perception and synthesis as distinct conditionings within a shared Generalist Open-World Temporal Perception Architecture (GOWTPA). Recognition, structured prediction, and simulation arise as different conditionings of the same generative substrate, while reasoning and embodiment-specific policies build upon the resulting world state. This positions generalist temporal perception as a potential foundation layer for broader multimodal intelligence and physical AI.

Large language models are transforming many industries with their text generation abilities. However, their outputs can be easily tampered with, creating serious risks in critical areas such as food safety reporting. To protect the integrity and traceability of AI-generated content, this paper introduces ToSS (Token Oriented Repartitioning and Strategic Selection), a reliable authentication method using adaptive dual watermarking. The key innovation of ToSS is its dual watermark encoding approach that divides vocabulary tokens into black and white sublists, enabling precise bit-level embedding of traceability information. Additionally, an entropy adaptive mechanism dynamically selects text regions with high prediction uncertainty for watermark insertion, maintaining text fluency and factual accuracy while ensuring reliable traceability. Experiments on multiple datasets, including food domain texts, demonstrate that ToSS achieves leading performance in both watermark capacity and decoding accuracy.

Large language models (LLMs) are often considered fragile under aggressive sparsification, and maintaining reliable performance typically requires sticking to moderate sparsity levels. However, recent studies suggest that LLMs are more resilient to high sparsity than previously thought, reframing the problem as a design challenge rather than a fundamental limitation. In this work, we challenge the perceived limits of unstructured post-training LLM pruning by revisiting elementary pruning strategies that have remained relatively underexplored at this scale. Through a progressive sparsification framework with second-order saliency and continued training coordinated with sparsity progression, we show that pretrained LLMs can retain strong performance far beyond commonly studied sparsity regimes. Across LLaMA-2 and Qwen-3 model families, our approach improves perplexity and downstream accuracy up to 99% sparsity, surpassing both the current state-of-the-art and representative baselines. Precisely, on LLaMA-2-7B, our approach achieves WikiText-2 perplexities of 13.48 and 19.67 at 95% and 99% sparsity, respectively, while delivering 3.23$\times$ decoding speedup and 6.21$\times$ memory savings at 95% sparsity. Taken together, our results show that LLMs can be pushed into extreme sparsity while retaining strong performance, providing a foundation for further improving sparse models in this regime.

The In-context learning (ICL) paradigm aids large language models (LLMs) to adapt to new tasks without need for fine-tuning. However, selecting an optimal combination of demonstration examples from a large pool of example subsets is a challenging problem. Existing approaches for selection do not model the complex relationship between ICL samples and downstream LLM performance. They typically perform static task-level selection, choosing subsets once offline, which can fail to generalize to unseen queries. We introduce DearICL (Data Efficient Algorithm for Ranking) ICL samples, a new framework that models demonstration example selection as a subset ranking problem. DearICL employs a non-linear surrogate employing a differentiable sorting objective within a gap-index bandit algorithm. The gap-index based approach enables fine-grained separation of good arms and borderline arms, which is used as an auxiliary objective to train the non-linear surrogate through sufficient sampling of borderline arms, supporting instance-level subset ranking. On exemplar selection benchmarks with open-source LLMs, DearICL achieves 8.08-15.9% accuracy gains over strong linear bandit baselines, with low sample complexity. Code and data: https://github.com/VenkteshV/DearICL.

Large language models (LLMs) are increasingly used or proposed for educational scoring, but single-model and single-run evaluations provide limited evidence for assessment use. Short-answer scoring requires evidence about reliability, validity, severity, diagnostic value, and failure cases. This study evaluated repeated multi-model OCG-PRES guided LLM scoring for short-answer assessment. The analysis used 996 SciEntsBank responses. GPT, DeepSeek, and Qianwen each scored every response across three independent runs using five OCG-PRES dimensions: concept coverage, relation accuracy, reasoning completeness, contradiction control, and domain relevance. Scores were evaluated against official binary and five-category labels and compared with non-LLM baselines based on answer length, Jaccard keyword overlap, TF-IDF cosine similarity, and a combined traditional logistic model. Repeated-run reliability was high for all models, with ICC(3,k) = .977 for GPT, .992 for DeepSeek, and .981 for Qianwen. DeepSeek was the most stable across runs. GPT showed the strongest official-label alignment by AUC (.909), while Qianwen was stricter, with higher precision but lower recall under the fixed threshold = 3.0 rule. OCG-PRES scores followed expected diagnostic patterns across five official categories and outperformed all non-LLM baselines in AUC and F1. Repeated multi-model OCG-PRES scoring provides reliability, validity, and diagnostic evidence for LLM-assisted short-answer scoring. The findings support cautious, evidence-based use as a scoring support tool rather than a replacement for human judgement.

Multi-agent reinforcement learning in networked populations is governed by the interaction between individual adaptation, local encounters, and changing environmental conditions. To study this interaction, we formulate a coupled learning--environment model in which agents update stateless $Q$-values on a fixed graph, while their population-average behavior drives an environmental variable that dynamically modifies the payoff matrix. Under a first-order mean-field closure, we derive a deterministic transport equation for the population distribution of $Q$-values and couple it with a projected discrete update for the environmental state. The resulting model is evaluated against finite-network Monte Carlo simulations on random regular, Erdős--Rényi, Barabási--Albert, and random geometric graphs. Across the tested parameter ranges, the mean-field system reproduces the main macroscopic cooperation and environmental trajectories, and the trajectory-level root-mean-square error generally decreases with population size and average degree. The analysis further shows that environmental feedback reshapes the learned action-value ordering, while reinforcing feedback can produce pronounced dependence on the initial learning bias and resource level. The environmental timescale also plays an important role: a rapid response can drive the resource state to a boundary before learning adapts, whereas a slower response preserves the interaction between behavioral learning and environmental recovery. These results provide a population-level description of coupled reinforcement learning and environmental dynamics and characterize the performance of the mean-field approximation within the tested network and parameter ranges.

Existing benchmarks for autonomous AI scientists evaluate only final outputs---generated code, hypotheses, or papers---yet discard the reasoning process by which those outputs were obtained. This makes it impossible to audit scientific methodology, diagnose failure modes, or distinguish systematic reasoning from fortunate guessing. We present OpenDiscoveryTrace, a public dataset of 558 complete AI scientific agent trajectories that captures how models reason, not just what they produce. Each trajectory records a structured 9-field-per-step trace---including thoughts, tool calls, observations, errors, revision triggers, and self-reported confidence---as models execute 124 scientific tasks spanning drug discovery, materials science, genomics, and scientific literature analysis. The dataset covers seven models: three frontier models (GPT-5.4, Claude Opus 4.6, and Gemini 3.1 Pro; 124 trajectories each, fully balanced across domains and difficulty levels) and four open-weight models (Qwen2.5-7B, Mistral-7B-v0.3, Phi-3.5-mini, and Qwen2.5-1.5B; 30 each), plus 60 live-retrieval variant trajectories. Pilot analysis on 363 LLM-judged trajectories reveals that process traces expose behavioral differences invisible to output-only evaluation: all three frontier models achieve comparable success rates (84--89%), yet Claude Opus 4.6 produces 30$\times$ more errors than GPT-5.4 (2.5 vs. 0.08 per trajectory, $p < 0.0001$, Cliff's $δ= 0.613$), with qualitatively different error profiles---66.7% tool misuse for Claude versus 83.6% reasoning errors for GPT-5.4. We define five benchmark tasks with baselines from logistic regression, random forests, LSTMs, and Transformer models. The dataset, trace schema, agent harness, and benchmark definitions are publicly available under CC BY 4.0 to support research on process-level evaluation, scientific agent auditing, and AI governance.

Power-of-two (PoT) quantisation turns a multiplication into a bit shift, so far only for the post-softmax attention--value product. The earlier and larger product, S=QK^T, has not been reformulated the same way. We quantise the key cache to a signed power-of-two fixed-point code, so that every scalar multiplication in QK^T becomes a sign flip, a bit shift and an integer accumulation. A fixed-point head-room F>= e_max turns every shift index r=F-e into a non-negative left shift, making the accumulation exact in integer arithmetic. We add a sub-power mantissa extension that halves the score error for one extra shift-add, a shift-exact online softmax whose running maximum lives in an integer log2 domain so that every rescaling is itself an exact shift, and a log-quantised AV product. Fused CUDA kernels in a 1.1B Llama decoder on an RTX 4090: at B=8, T=32k the 4-bit kernel is 4.60x faster than FP16 scaled dot-product attention with a 2.5x smaller KV cache, and batched decoding crosses over at B=64 to 1.22x FP16 throughput at 11.02 against 13.77 GB peak memory. A matched INT8 multiply--accumulate kernel on the same tiling reaches 5.29x: on a GPU with a hardware 4-way INT8 dot product, shift-accumulate is not faster than the MAC it replaces. An iso-storage control--an 8-bit shift code of exactly one byte per key, run through the identical kernel--prices the arithmetic substitution alone at 2.02x at comparable accuracy. The advantage of the PoT representation is therefore its density and the removal of the multiplier, not raw GPU throughput. A sweep over exponent width, mantissa levels, granularity and rounding gives the design lesson: nearest-PoT relative error is scale-free, so extra exponent bits buy nothing (eps_S=0.1302 at 3, 4 and 5 bits) and accuracy must come from mantissa terms.

Guided depth completion methods heavily depend on RGB quality and alignment, while unguided ones often suffer from limited precision due to the absence of explicit visual cues. In this paper, we present Depth-to-Image Synthesis-Driven Generative Unguided Depth Completion (GUDC), a new completion paradigm that innovatively bridges advanced 2D generative models with unguided depth completion, enabling semantics-aware depth inference without real RGB inputs. Our key idea is to exploit ControlNet's powerful depth-conditioned generation capability to synthesize pseudo-images directly from sparse depth, effectively converting the original unguided setting into a semantics-guided one. To address the potential image-depth misalignment caused by depth sparsity, we propose a multi-level dense-to-sparse representation distillation strategy for ControlNet fine-tuning, where dense-depth features act as teacher signals to distill consistent structural representations for sparse-depth inputs. Furthermore, during pseudo-image-guided completion, we propose a pseudo-image semantic attention fusion module to adaptively extract informative semantic cues from pseudo-images while suppressing artifacts (e.g., texture hallucinations). Extensive experiments on KITTI and NYUv2 validate that our GUDC achieves superior accuracy and robustness over existing methods.

Verifiable outcome rewards guide language-model post-training, but sequence-level advantages do not identify which token-level decisions should be preserved or revised. Evidence-conditioned Teachers provide denser supervision by replaying sampled trajectories with privileged feedback. Yet indiscriminate imitation risks transferring formatting or reasoning-style shifts that do not support task success. We introduce VERPO, a Verified Evidence Regularized Policy Optimization framework that treats evidence as a proposal for policy correction while retaining the outcome objective. It separates evidence-free reference restoration from signed token-level evidence corrections. Fisher Evidence Contrast attenuates corrections along an estimated evidence-presence direction. A stopped token-wise ZPD controller scales acceptance according to local reward alignment and Fisher movement cost, while the reference channel remains independent of acceptance. Across five scientific-reasoning and tool-use tasks, the best variant on each backbone exceeds the strongest compared baseline in average score. The averages rise from 0.6826 to 0.6857 on Qwen3-4B, from 0.6895 to 0.7058 on Qwen3-8B, and from 0.4751 to 0.5657 on Llama-3.2-1B.

Activation steering controls LLM behavior at inference time by adding learned directions to hidden states, but existing methods handle one concept at a time. Pluralistic alignment, where different stakeholders need different value emphases, requires steering multiple dimensions simultaneously. We show that naive steering produces substantial spillover: the effect intended for one value leaks into others. This parallels the treatment-versus-spillover decomposition in causal inference. We trace spillover to geometric entanglement of steering directions, captured by their Gram matrix, and derive a zero-cost correction from an activation-norm-penalized objective that decouples each direction's contribution exactly. Our end-to-end pipeline requires no fine-tuning, no reward model, and no manual prompt engineering: given only domain questions, it automatically discovers value dimensions, extracts directions, diagnoses entanglement, and applies corrected steering. On climate discourse, the correction improves the net steering effect from +5.9% to +14.0%, validated over 100,000 pairwise judgments.

Reinforcement learning (RL) excels on tasks with verifiable rewards, but in open-ended tasks, the reliability of reward models remains a key challenge. Existing solutions either depend on costly proprietary LLM-as-a-Judge systems or opaque scalar reward models that lack interpretability. Recent works on generative reward models offer a promising alternative, but they remain constrained by static evaluation criteria, fragmented evaluation paradigms, and limited multilingual support. To address these challenges, we introduce MixReward, a large-scale multilingual dataset spanning six domains and 103 languages, containing both pairwise and listwise data, and propose UniRRM, a unified reasoning reward model supporting multiple languages and evaluation paradigms. UniRRM uses a staged reasoning chain to dynamically generate task-generic and instruction-specific criteria, enabling fine-grained, input-adaptive judgments while maintaining consistency across languages. Experiments demonstrate that UniRRM-8B and UniRRM-14B achieve performance close to the state-of-the-art for models of comparable size across multiple benchmarks, and are effective for unseen evaluation paradigms. In addition, ablation studies validate the reliability and effectiveness of UniRRM.

As machine translation (MT) systems continue to improve, standard benchmarks become less informative for exposing remaining weaknesses. Traditional methods for creating challenging test sets rely on expensive manual creation or curation, while automated approaches struggle to produce sets with the necessary translation difficulty and linguistic diversity. We propose a scalable reinforcement-learning-based approach for rewriting existing source texts into instances that are more difficult to translate for MT systems. We fine-tune a large language model with Group Relative Policy Optimization (GRPO), using reward signals based on translation difficulty together with constraints for semantic similarity, grammaticality, and approximate length preservation. On WMT25, our approach substantially reduces average COMET translation quality from 0.63 to 0.48, while preserving grammaticality and readability, whereas the base model remains at 0.64. Evaluations on the unseen WMT19-WMT24 benchmarks confirm that this behavior generalizes beyond the training data, and human evaluation further shows that the rewrites substantially lower translation quality while incurring a moderate drop in naturalness and only a small change in grammaticality. We release our code to support reproducibility.

Federated fine-tuning adapts large language models (LLMs) to decentralized client data, but its scalability in cross-device training is often limited by the high communication cost. Muon is an optimizer that improves optimization performance by orthogonalizing momentum for matrix-valued parameters. Existing federated Muon methods demonstrate the benefit of matrix-aware optimization in federated learning, but still require transmitting full layer-size updates and optimizer state. A natural way to reduce communication is to directly apply Muon to LoRA factors, but this changes the optimized object and weakens Muon's matrix-aware update geometry. We propose FedSubMuon, a communication-efficient federated Muon fine-tuning method that optimizes compact coefficient matrices within shared structured subspaces. This design keeps Muon on a single matrix-valued trainable object, while reducing the client upload to compact coefficient matrices. We further introduce FedSubMuon-GT, an accuracy-oriented extension that uses projected gradients to adapt tracked subspace bases toward task-relevant gradient directions. Experiments on instruction tuning and mathematical reasoning show that FedSubMuon-GT achieves the best overall accuracy on four of five dataset-model pairs, while FedSubMuon performs best under all matched communication budgets. On Dolly-15K, the closest communication baseline requires 5.5 times and 1.4 times more total communication on Llama-1B and Qwen-4B, respectively.

Large language models answering questions over multi-page documents are expected to cite the supporting pages, yet supplied citations are sometimes inaccurate, and current evaluations score citations at generation time or against text passages: no existing benchmark evaluates whether a system can verify and correct a page-level citation already attached to an answer. We propose AtomCite, an agentic framework that parses an answer into claims, checks each claim against the image of its cited page, and applies a deterministic repair policy. To evaluate it, we introduce DocCite, to our knowledge the first benchmark for systems that verify and correct page-level citations in document images. Built on MP-DocVQA and DUDE, it combines 928 validated injected instances with 2,468 candidate natural errors harvested from frontier- and efficiency-tier models, of which a two-annotator audit confirms 1,909 as genuine errors. Primary labels are assigned deterministically, not by LLM judges, with the human audit as a separate validation layer. Across three model families (Gemini, Claude, and GPT), AtomCite reaches around 93% binary verification accuracy on the injected benchmark, significantly outperforming every OCR-only condition, including a compute-matched control, and exceeding every prior text-based baseline given the same OCR text. Its repair policy lifts citation precision on the injected mix from a constructed 34% to 87-90% while retaining over 90% of correct claims. AtomCite also transfers: with frozen prompts and zero training, it raises the hallucination-detection scores of two open 7-8B models on five public benchmarks above the same models prompted as direct judges. Finally, the audit shows that noise in automatic labels biases measured verifier accuracy and can reverse system rankings, so evaluations relying only on synthetic or automatic labels risk mismeasuring verification capability.

Multi-fingered dexterous manipulation remains a frontier for real-world reinforcement learning (RL) due to the high-dimensional action space and the prohibitive cost of hardware failures. While human-in-the-loop (HIL) RL allows operators to intervene before failures occur, current pipelines often treat these interventions as reactive corrections, discarding the rich safety signal inherent in the operator's decision to take control. In this paper, we ask: How can we learn from what a human would avoid? We present WHIRL, a safety-aware RL framework that transforms binary human interventions into forward-predictive signals for proactive risk avoidance. Our approach centers on an intervention-aware latent world model with four prediction heads: dynamics, reward, termination, and a novel per-state intervention-probability head that learns to predict the likelihood of a human takeover at future states. This head provides an actor-side risk-shaping term that discourages the policy from entering "intervention-prone" regions, modeling the operator's internal safety threshold. We evaluate our framework on a 16-DoF LEAP Hand across tasks spanning convex and irregular object grasping, prismatic manipulation, and long-horizon multi-stage tasks. Our results show that predictive risk-shaping enables the system to achieve a 96.7 percent success rate on complex grasping tasks while reducing the operator intervention burden by up to 84 percent in step-weighted terms. By closing the loop between human intuition and predictive world modeling, this work provides a practical safety-aware recipe for training complex dexterous agents in the real world while reducing operator fatigue and hardware-risk exposure.

Reconstructing 4D scenes from casually captured monocular video is vital for applications in virtual reality (VR) and embodied AI. Recent advances in 4D reconstruction and novel view synthesis have substantially propelled this capability. However, existing reconstruction methods generally cannot recover regions beyond visible camera limits. Consequently, we introduce a new paradigm that achieves 4D scene synthesis by combining visible-region reconstruction from monocular input with invisible-region generation beyond observable camera boundaries. We present Panoramic Alignment for Strategic Exploitation of Generative Priors (PASTEL). Specifically, PASTEL proposes panoramic scene alignment, a novel representation that reformulates the intractable 3D "invisible region" exploration into a tractable 2D directional trajectory planning. This is achieved by reducing the viewpoint planning from 6-DoF search to a 2D directional search with explicit visibility boundaries. By operating within this panoramic space, our method strategically identifies camera trajectories that maximize exploration beyond observable boundaries while minimizing viewpoint deviation. Experimental results show that PASTEL can not only extrapolate plausible scene content beyond the observable boundaries of input monocular videos, but also substantially boost monocular 4D reconstruction performance. PASTEL outperforms the previous state-of-the-art method by 0.9dB in full-image PSNR on the DyCheck IPhone dataset.

Diffusion Transformers have become the dominant paradigm in generative AI, but their high computational costs severely hinder real-time applications. Prediction-based feature caching is widely used to accelerate diffusion transformers; however, as the number of steps increases, the deviation between its predictions and the reference full-compute trajectory gradually grows. An intuitive idea is to use an online regression model to dynamically correct this deviation, but it faces the issue of label data being unavailable during the acceleration process. This paper presents a statistical observation that the residuals between the features of full computation steps using caching methods and reference full-compute trajectory locally exhibit a zero-mean Gaussian distribution. By treating the features of full computation steps as noisy observations of reference features, the data acquisition problem is resolved. Based on this observation, a plug-and-play GP-Refiner correction framework is proposed. This method utilizes Gaussian Process Regression for correction and, leveraging the properties of GPR, introduces an uncertainty-adaptive computation strategy that triggers necessary full-computation calibration by monitoring the posterior variance in real time. Experiments demonstrate significant improvements across different models when combined with various state-of-the-art methods. Integrating the proposed framework with TaylorSeer reduces the computational load by 19.3% while improving PSNR by 0.9 dB and reducing LPIPS from 0.46 to 0.29. Code is available in https://github.com/Aredstone/GP-Refiner.

Red-teaming Text-to-Image (T2I) models is essential for safe deployment, yet it remains particularly challenging against implicit adversarial prompts. Unlike explicit adversarial prompts that can be readily identified and blocked, implicit ones are much harder to detect: the prompts appear benign on the text surface yet still lead to inappropriate visual content. To address this, we propose Adversarial Probing for Implicit VulnErabilities (AdvPIE), a multimodal agentic framework to expose implicit vulnerabilities without requiring access to the parameters of target models. AdvPIE adopts a policy agent to generate and refine implicit adversarial prompts based on the feedback from a judge agent. To construct informative feedback, the judge agent provides modality-specific safety evaluation at both global and relative levels across iterations. To effectively leverage the feedback, we propose a novel Cumulative Adversarial Decoding strategy for the policy agent, which dynamically reweights token distributions to favor tokens that lead to more harmful images while preserving sampling diversity. Extensive experiments on standard and safety-aligned T2I models show that AdvPIE1 effectively uncovers implicit vulnerabilities, outperforming various baseline methods.

Vision foundation models trained on Internet-scale RGB datasets enable remarkable capabilities across a range of tasks, from text-to-video generation to few-shot 3D scene reconstruction. An acoustic foundation model trained on large-scale sonar datasets could enable similar capabilities in the underwater domain, where turbidity and low-visibility conditions make conventional RGB foundation models inapplicable. Unfortunately, a lack of freely available large-scale sonar datasets makes training such a model from scratch impractical. In this work, we demonstrate that vision foundation models can be efficiently adapted to the sonar setting by (1) exploiting the geometric relationship between the two sensing modalities and (2) employing accurate physics-based noise models for synthetic data generation. The resulting sonar adaptation models enable new capabilities: For the first time, we experimentally demonstrate sonar-based pose-free 3D reconstruction.

World models promise a general route to embodied intelligence: learn predictive dynamics once, then reason, plan, and act with them. Increasingly, the representations beneath such models are pretrained on large-scale video, interaction, and multimodal corpora, which raises a question prediction quality alone cannot answer: when is a learned representation actually actionable? We identify a failure mode we call counterfactual collapse: a model predicts visually plausible futures while failing to distinguish interventions with different behavioral consequences. This arises whenever a representation is optimized for perceptual similarity rather than intervention structure, which is precisely the objective under which most large-scale pretrained encoders are learned. We introduce Counterfactual Latent World Models (CLWM), which combine a recurrent belief-state encoder, action-conditioned latent dynamics, and a contrastive counterfactual objective that separates futures induced by distinct interventions even when their observations look alike. Across occluded manipulation, aliased navigation, and long-horizon manipulation, CLWM improves planning success over the strongest baseline (65.1% $\to$ 74.6% on Occluded Push and 67.3% $\to$ 78.9% on Aliased Maze) and reduces exploitative planning failures (18.4% $\to$ 9.7% on Deferred Kitchen), with ablations attributing the gains to hard counterfactual negatives, especially perceptual-alias negatives. Finally, our counterfactual separability metric, which tracks planning success across the five baseline model classes ($r \ge 0.94$), is representation-agnostic: given intervention-outcome labels, it can audit any encoder, pretrained or trained from scratch, before a planner trusts it. We do not yet measure it on large-scale pretrained encoders. Here we establish the metric and its relationship to planning success for world models trained from scratch.

Aligning a text-to-image generation flow model with a reward makes it follow objectives that the training data alone does not provide. Alignment fine-tuning delivers this by reinforcement learning (RL) or preference optimization, but it must be repeated for every checkpoint and returns a model fixed at the reward and strength it was trained with. Test-time alignment instead steers a frozen model during sampling, allowing task-specific and sample-specific guidance. Existing methods obtain this only by drawing the per-step signal from the reward function itself, through its gradient, or through a separately trained value function. We propose changing the supervision source: let a pair of weak models, not a reward function, supply the supervision. A source aligned model, kept together with its base as a source alignment pair, stores its training reward as an implicit, step-wise, KL-anchored signal expressed in the sampler's own coordinates. We explore whether this model-form supervision can cross scale, and show that it does: our method, AlignGraft, aligns a larger, frozen, never-tuned model by adding the pair's velocity difference during sampling. The transport is exact under a shared noising kernel and needs neither the reward nor its gradient at test time. The method has no schedules, only a single scalar that controls the alignment strength and can extrapolate it beyond that of the source alignment pair. Across image and video flow models (Stable Diffusion 3.5, FLUX, and Wan), the transfer lifts the frozen large model on preference, compositional, and text-rendering rewards, can exceed the source aligned model itself, and preserves the large model's fidelity at a small constant sampling overhead. Extensive experiments show that one alignment run on a weak model produces supervision that the whole model family can reuse at test time.

Recent image generation systems increasingly combine multimodal understanding, reasoning, and synthesis, suggesting that they may do more than render plausible scenes. Yet existing evaluations emphasize aesthetics, prompt alignment, compositionality, or text-based answers, leaving unclear whether these systems can solve visual problems and faithfully express solutions in pixels. We introduce GenPuzzle, a benchmark for reasoning-centric image generation. GenPuzzle contains 2,005 problems across 12 tracks, spanning pattern completion, spatial construction, mazes, Sudoku, nonograms, tangrams, board games, matchstick puzzles, orthographic projection, and mathematical visual proof. Each task provides a visual puzzle and requires an image output that preserves the input state while executing a logically valid solution. GenPuzzle uses task-specific evaluation protocols: discrete grid outputs are transcribed and verified programmatically, while visually complex outputs are assessed with tiered, multidimensional, or binary multimodal large language model (MLLM) rubrics. We further select the automatic judge by measuring agreement with human reference scores. Across three frontier generators, the strongest model reaches only 40.57 Macro Overall, revealing frequent failures in logic, geometry, state preservation, and instruction execution. GenPuzzle provides a testbed for measuring progress from image rendering toward visual problem solving.

We present PAI-Actor, a cinematic multi-character animation framework for character replacement in dynamic movie scenes. Unlike conventional animation systems that mainly drive a single static image or a single subject, our goal is to replace and animate multiple characters within real video clips while preserving the original scene dynamics, camera motion, and background content. This setting is particularly challenging because the generated characters must remain consistent with the source performance in motion and interaction, while also matching the surrounding background in lighting, shadow, composition, and overall cinematic appearance. To address this, we formulate multi-character animation as a structure-guided human recovery problem and build a movie-driven training pipeline from high-quality film data. Furthermore, to support practical cinematic production, we introduce a bidirectional-to-autoregressive distillation framework: we first train a bidirectional diffusion transformer for high-quality short-clip generation at 1080P resolution, and then distill it into an autoregressive video-to-video model for efficient inference and longer video generation. Experiments show that PAI-Actor enables high-fidelity multi-character animation with strong scene consistency, cinematic visual quality, and efficient long-form generation.

Image-to-video generators are often credited with absorbing physical dynamics as implicit world models, a claim the community currently checks with plausibility scores that ask whether a clip looks consistent with real-world motion. Plausibility is the wrong test on its own, because a clip can look natural while encoding the wrong value of the governing physical parameter, and no existing benchmark measures this gap directly. PhysWeep closes it with a fixed, label-free audit, treating a frozen generator as a black box, recovering the realized parameter from generated pixels, and reporting how often generation is trackable at all, how far the realized value sits from the requested one, and which, if either, of the literature's two proposed failure mechanisms the data support. A deterministic-simulator positive control confirms every score is exactly checkable. Applied to three open generators across six sweep axes, PhysWeep finds a specific, reproducible, previously undocumented failure. Conditional on producing trackable motion, two of the three generate confident, well-fit dynamics that converge to one of a small number of fixed, wrong values selected by the sampling seed rather than by the request, reproducing across two independent model families, two physical systems, and an independent tracker. It matches neither the prior reversion nor the case-based clamping the literature anticipates, because the reversion target is seed-conditional rather than a single global default, and a leave-one-out selection rule rejects both; the in-range faithfulness slope is statistically indistinguishable from zero wherever a response is estimable at all. A benchmark averaging over seeds would never see this: each sample is confidently locked to a wrong constant, exactly the failure a plausibility score is structurally blind to. We release the protocol, suite, and analysis code.