Search arXiv⌕ Search

arXiv subjects

Yizhe Zhang

Publications and source records attributed to Yizhe Zhang.

At least 19 recordsLinked to original sources
arXiv2026

Event-STVG: Spatio-Temporal Video Grounding for Long-Form Event Reasoning

Spatio-Temporal Video Grounding (STVG) aims to jointly localize a target in space and time given a natural-language query. Despite continued advances, existing STVG benchmarks remain largely centered on localizing targets through instance-specific descriptions, leaving them largely decoupled from long-form reasoning increasingly emphasized in multimodal large language models (MLLMs). In this paper, we introduce Event-STVG, a task that shifts STVG from locating what is described toward event-level reasoning by unifying long-form context, repeated-event reasoning, event localization, role inference, and multi-target grounding. To facilitate research on Event-STVG, we construct a large-scale benchmark comprising 4,016 annotated event occurrences across soccer and seven additional domains. Experiments with state-of-the-art MLLMs reveal substantial difficulty, with the strongest model achieving only 5.4% m-vIoU at the 10-minute horizon. Across domains, we consistently observe confusion among repeated same-type events, spatial grounding bottlenecks, and rare success in jointly grounding all required roles. Event-STVG establishes a new challenge for visual understanding in MLLMs, calling for future models that can jointly reason about complex events across space and time. Code and annotations are available at https://github.com/THU-KEG/EventSTVG

cs.CV↗
arXiv2026

HarnessTTT: Learn to Optimize Harness at Test-Time for Scientific Discovery

Scientific discovery with language models requires finding the best valid solution within a finite computational budget. We study how feedback from this search can improve the procedure used to conduct it. We introduce HarnessTTT, a test-time learning framework that trains a harness proposer while keeping the executor frozen. The proposer generates executable harnesses comprising system prompts, reusable skills, tools, middleware, and execution controls. Each harness guides a budgeted discovery rollout, whose outcome provides feedback for updating the proposer. A discovery-oriented policy-gradient objective combines gap-normalized rewards, leave-one-out advantages (RLOO), and max-sharpening to favor high-performing proposals. Generated components are validated before execution, and the search retains the best valid solution and selected harnesses across rounds. Across fourteen discovery tasks in mathematics, systems, GPU kernels, heuristic programming, astrodynamics, and biology, HarnessTTT with a frozen 9B executor improves on OpenEvolve with the same executor on all eleven benchmark tasks (e.g., from 1.17 to 2.64 on circle packing), improves on the discovery-fine-tuned Finch-9B by up to 92%, holds the best open-weight sub-10B result on nine, and reaches 0.380919 on Erdős minimum overlap, better than the best human reference of 0.380927. Scaling the proposer and frozen executor to 27B improves on eleven of fourteen tasks, and transferring the learned harnesses to GPT-5.5 matches or exceeds the strongest reported system of any backbone on nine of fourteen tasks. Together, these results show that harness search can itself be learned at test time: a trained proposer lets a small frozen executor search effectively, and the harnesses it finds carry over to stronger executors.

cs.LG↗
arXiv2026

SkillFM: Generating Skills for LLM Agents via Latent Flow Matching

Textual skills provide reusable guidance for large language model agents, but existing approaches often rely on manually curated skill banks or reinforcement learning with indirect and delayed feedback. We introduce SkillFM (Skill Flow Matching), a generative framework that synthesizes task-conditioned textual skills directly without test-time skill retrieval. Our framework combines a codec for encoding and reconstructing textual skills in a continuous latent space with a conditional flow model trained using improved MeanFlow. At inference time, the learned velocity field enables single-step latent sampling, and an LLM-based decoder converts the sampled representation into textual guidance for a frozen downstream agent. We evaluate the framework on embodied tasks, question answering, and web shopping. On ALFWorld and Search-QA, our method achieves the best overall performance among the compared vector-based skill approaches. Our analyses further demonstrate that latent skill generation is an effective alternative to retrieval-based skill augmentation. Our code and training skill libraries are available at https://github.com/lulushang999/SkillFM.

cs.AI↗
arXiv2026

Uni-LaDiR: Latent Diffusion Unifies Multimodal Reasoning

Multimodal models increasingly think with different modalities such as images, 3D point clouds, and robot states, not just text. Yet each modality is still encoded into its own representation space, creating a modality-switching gap whenever reasoning moves from one modality to another. In this paper, we introduce Uni-LaDiR (Unified Latent Diffusion Reasoner), a framework that unifies different modalities into a shared latent space for multimodal reasoning. A unified encoder maps teacher reasoning steps from different modalities into latent thought tokens in a shared space, trained to extract the information needed for later reasoning steps and the final output. A diffusion reasoner, trained jointly with the encoder, generates these tokens at inference without teacher reasoning steps. Across eleven vision-language model (VLM) benchmarks and two vision-language-action (VLA) suites, Uni-LaDiR achieves relative gains over the strongest baselines of 7.3% on four mathematical and logical VLM benchmarks and 6.1% on RLBench manipulation tasks. Controlled comparisons show increasing gains as more teacher modalities are unified. These results suggest that unification improves multimodal reasoning by weaving it into a single thread, where the model predicts successive thoughts in a common representation space.

cs.LG↗
arXiv2026

Unmasking On-Policy Distillation: Where It Helps, Where It Hurts, and Why

On-policy distillation offers dense, per-token supervision for training reasoning models; however, it remains unclear under which conditions this signal is beneficial and under which it is detrimental. Which teacher model should be used, and in the case of self-distillation, which specific context should serve as the supervisory signal? Does the optimal choice vary from one token to the next? At present, addressing these questions typically requires costly training runs whose aggregate performance metrics obscure the dynamics at the level of individual tokens. We introduce a training-free diagnostic framework that operates at the highest resolution: per token, per question, and per teacher. We derive an ideal per-node gradient defined as the parameter update that maximally increases the student's probability of success. We then develop a scalable targeted-rollout algorithm to estimate this gradient efficiently, even for long chains of intermediate thoughts. The gradient alignment score, defined as the cosine similarity between this ideal gradient and any given distillation gradient, quantifies the extent to which a particular configuration approximates the ideal signal. Across a range of self-distillation settings and external teacher models, we observe that distillation guidance exhibits substantially higher alignment with the ideal on incorrect rollouts than on correct ones, where the student already performs well and the teacher's signal tends to become noisy. Furthermore, we find that the optimal distillation context depends jointly on the student model's capacity and the target task, and that no single universally effective configuration emerges. These findings motivate the use of per-task, per-token diagnostic analyses for distillation.

cs.LG↗
arXiv2026

DeepWeaver: Bridging the Evidence Synthesis Gap in Open-Ended Question Answering

Retrieve-then-generate pipelines are commonly used to produce deep-research answers for open-ended questions, but retrieval alone is insufficient: LLMs must organize noisy and fragmented evidence into comprehensive, well-cited answers. We refer to this process as evidence synthesis. However, direct generation often underuses evidence, misaligns citations, and collapses diverse information into shallow summaries, exposing an evidence synthesis gap between retrieval and generation. Thus, we propose DeepWeaver, a novel framework that weaves noisy retrieved evidence into comprehensive answers by maintaining Thought Block Chains (TBCs), a structured representation that groups claims, salient information, keywords, and supporting evidence. DeepWeaver uses subordinate TBCs to inspect residual evidence, commit TBC revisions, and discover new claims before final generation. We evaluate DeepWeaver on open-ended QA over both knowledge bases and the web, and introduce LoQA, a high-density benchmark for evidence synthesis. Across multiple LLMs, DeepWeaver improves content sufficiency, citation grounding, and detail preservation on LoQA, while achieving deeper insights and higher citation quality on DeepResearch Bench. These results show that evidence weaving is an effective mechanism for bridging retrieval and generation in open-ended QA. Our code is available at https://github.com/KlozeWang/DeepWeaver.

cs.CL↗
arXiv2026

SimpleDesign: A Joint Model for Protein Sequence and Structure Codesign

Proteins are fundamental to biological processes, with their function determined by the complex interplay between the amino acid sequence and the three-dimensional structure. Developing generative models capable of understanding this intrinsically multi-modal relationship is crucial for fields like drug discovery and protein engineering. Existing models often rely on a multi-stage training process where autoencoders that tokenize data into latent representations are trained in a first stage. Secondly, a generative model is trained on the latent representation of the autoencoder(s), i.e., generative modeling in a latent space. We hypothesize that this multi-stage training is not necessary to obtain performant co-design models and thus present SimpleDesign, an effective multi-modal protein design model trained directly in the data space. SimpleDesign leverages a single-stage end-to-end objective that combines discrete cross-entropy for sequences and a regression objective for structures. In order to effectively model the difference in sequence and structure modalities, we develop a Mixture-of-Transformer architecture that allows modality-specific processing while keeping global self-attention over both modalities. We train SimpleDesign on over 2M sequence-structure pairs achieving strong performance across co-design and unconditional sequence/structure generation benchmarks.

cs.LG↗
arXiv2026

Your Harness is Not Secure: Benchmarking Real-world Threat of Command Line Interface Agent

Command-line interface (CLI) agents powered by large language models (LLMs) can interpret natural-language requests, plan multi-step tasks, execute shell commands, and modify files and system state. As these agents are increasingly used for operating-system (OS) workflows, it is important to evaluate whether they can be misused to carry out security-relevant operations. Existing benchmarks often lack an attacker-knowledge model grounded in tactics, techniques, and procedures (TTPs), provide limited coverage of end-to-end kill chains, rely on simplified single-host environments, or use LLM-as-a-judge for success evaluation. We introduce AdvCLI, an MITRE ATT&CK-aligned benchmark for evaluating OS-level misuse risks of CLI agents in a controlled multi-host sandbox. AdvCLI contains 140 tasks: 40 direct malicious requests, 74 TTP-based tasks, and 26 end-to-end kill chains. Each task is paired with deterministic hard-coded verification protocols that check whether the requested OS-level effect is realized. We evaluate seven CLI agents and products built on nine foundation models, including ReAct, OpenClaw, OpenAI Agent SDK, Claude Code, Gemini CLI, Cursor CLI, and Cursor IDE. Results show that current CLI agents frequently proceed beyond refusal and can complete a non-negligible fraction of malicious OS-level tasks, especially when requests include TTP-style attacker knowledge. AdvCLI provides a reproducible testbed for evaluating these risks and for developing stronger safety mechanisms for tool-using CLI agents in the future.

cs.CR↗
arXiv2026

Spotting (and Missing) Algorithmic Bias: Investigating User Understanding in a Fairness Assessment Tool

Fairness metric selection is typically left to data scientists, but which biases are problematic and which metric captures them best depends on stakeholders' experience and domain knowledge. This calls for involving non-technical stakeholders, but the research prototypes built for this purpose so far have not tested whether these stakeholders form accurate mental models of the metrics they interact with or can act on them to identify biases. We present FairAware, a fairness assessment tool co-designed with Human Resources (HR) domain experts. We evaluate stakeholders' understanding through a mixed-methods study with 70 participants (35 HR employees, 35 job seekers), measuring objective and subjective understanding, cognitive load, bias identification accuracy, and open-ended feedback. Most participants correctly identified the most disadvantaged group, with task duration being the only significant predictor. We also found a gap between subjective and objective understanding, with both groups performing similarly across all measures. These results suggest that fairness assessment tools for non-experts are usable for identifying biases but need built-in checks on understanding before stakeholders make higher-stakes decisions.

cs.HC↗
arXiv2026

ESRVS: Extreme Semi-Supervised Retinal Vessel Segmentation with a Single Annotated Image

Learning from minimal human supervision is a long-standing goal in medical image analysis, where dense expert annotations are costly. We study retinal vessel segmentation in an extreme semi-supervised setting with one annotated image and a pool of unlabeled images. We propose ESRVS, which selects a representative reference image for manual annotation and transfers vessel cues using target-domain-adapted DINOv3 features. ESRVS constructs a multi granular vessel prototype, combines prototype-similarity maps with a physics-inspired prior to generate initial pseudo-labels, and refines the transferred supervision through weighted pseudo-label training and adversarial refinement. Across eight public datasets, ESRVS achieves the best Dice and clDice on six datasets, and the best HD95 on all eight datasets among the compared semi-supervised methods, although those methods use 10 to 20% labeled data. With Mask2Former, ESRVS retains on average 93.7% of fully supervised Dice and 95.1% of fully supervised clDice. These results demonstrate the potential of foundation-model label propagation for highly label-efficient retinal vessel segmentation. Code is available at https://github.com/IAANNH/ESRVS.

cs.CV↗
arXiv2026

Spatial Lifting for Dense Prediction

We present Spatial Lifting (SL), a novel methodology for dense prediction tasks. SL operates by lifting standard inputs, such as 2D images, into a higher-dimensional space and subsequently processing them using networks designed for that higher dimension, such as a 3D U-Net. Counterintuitively, this dimensionality lifting allows us to achieve good performance on benchmark tasks compared to conventional approaches, while reducing inference costs and \textbf{drastically lowering the number of model parameters}. The SL framework produces intrinsically structured outputs along the lifted dimension. This emergent structure facilitates dense supervision during training and enables single-forward-pass self-consistency-based quality and uncertainty estimation at test time. Spatial Lifting introduces a simple and general modeling strategy that offers a promising path toward more efficient, accurate, and reliable deep networks for dense prediction tasks in vision.

cs.CV↗
arXiv2026

ChipChat: Low-Latency Cascaded Conversational Agent in MLX

The emergence of large language models (LLMs) has transformed spoken dialog systems, yet the optimal architecture for real-time on-device voice agents remains an open question. While end-to-end approaches promise theoretical advantages, cascaded systems (CSs) continue to outperform them in language understanding tasks, despite being constrained by sequential processing latency. In this work, we introduce ChipChat, a novel low-latency CS that overcomes traditional bottlenecks through architectural innovations and streaming optimizations. Our system integrates streaming (a) conversational speech recognition with mixture-of-experts, (b) state-action augmented LLM, (c) text-to-speech synthesis, (d) neural vocoder, and (e) speaker modeling. Implemented using MLX, ChipChat achieves sub-second response latency on a Mac Studio without dedicated GPUs, while preserving user privacy through complete on-device processing. Our work shows that strategically redesigned CSs can overcome their historical latency limitations, offering a promising path forward for practical voice-based AI agents.

eess.AS↗
arXiv2026

Embarrassingly Simple Self-Distillation Improves Code Generation

Can a large language model (LLM) improve at code generation using only its own raw outputs, without a verifier, a teacher model, or reinforcement learning? We answer in the affirmative with simple self-distillation (SSD): sample solutions from the model with certain temperature and truncation configurations, then fine-tune on those samples with standard supervised fine-tuning. SSD improves Qwen3-30B-Instruct from 42.4% to 55.3% pass@1 on LiveCodeBench v6, with gains concentrating on harder problems, and it generalizes across Qwen and Llama models at 4B, 8B, and 30B scale, including both instruct and thinking variants. To understand why such a simple method can work, we trace these gains to a precision-exploration conflict in LLM decoding and show that SSD reshapes token distributions in a context-dependent way, suppressing distractor tails where precision matters while preserving useful diversity where exploration matters. Taken together, SSD offers a complementary post-training direction for improving LLM code generation. Our code is available at https://github.com/apple/ml-ssd

cs.CL↗
arXiv2026

Latent Reasoning with Normalizing Flows

Large language models often improve reasoning by generating explicit chain-of-thought (CoT), demonstrating the importance of intermediate computation. However, textual CoT forces this computation through a discrete, serial, and communication-oriented token stream: each reasoning step must be verbalized before the model can proceed, even when the underlying update is semantic, uncertain, or only partially formed. Latent reasoning offers a higher-bandwidth alternative by performing intermediate computation in compact continuous states before committing to text. Yet existing latent-reasoning methods often sacrifice key advantages that make CoT effective in autoregressive language models, including native left-to-right generation, probabilistic sampling, compatibility with KV-cache decoding, and tractable likelihood estimation. We propose NF-CoT, a latent reasoning framework that preserves these advantages by modeling continuous thoughts with normalizing flows. NF-CoT instantiates a TARFlow-style normalizing flow inside the LLM backbone, defining a tractable probability model over compact continuous thoughts distilled from explicit CoT. Continuous-thought positions are generated by an NF head, while text positions are generated by the standard LM head within the same causal stream. This design provides exact likelihoods for latent thoughts, enables probabilistic left-to-right decoding with the original KV cache, and supports direct policy-gradient optimization in the latent reasoning space. On code-generation benchmarks, NF-CoT improves pass rates over explicit-CoT and prior latent-reasoning baselines while substantially reducing intermediate-reasoning cost.

cs.CL↗
arXiv2026

LT2: Linear-Time Looped Transformers

Looped Transformers (LT) have emerged as a powerful architecture by iterating their layers multiple times before decoding the final token. However, pairing them with full attention retains quadratic complexity, making them computationally expensive and slow. We introduce LT2 (Linear-Time Looped Transformers), a family of looped architectures that replace quadratic softmax attention with subquadratic, linear-time attention. We study two variants: LT2-linear with linear attention and LT2-sparse with sparse attention. We find that looping uniquely synergizes with these variants: it enables iterative memory refinement in linear attention and progressively expands the effective receptive field in sparse attention. We formalize these benefits theoretically and demonstrate consistent empirical gains across controlled recall, state-tracking, and language modeling tasks. We then explore LT2-hybrid, which combines different attention variants in a looped setting. Two variants are especially promising: LT2-hybrid (GDN+DSA), which interleaves linear and sparse attention to maximize efficiency and matches the standard looped transformer's quality at fully linear-time cost; and LT2-hybrid (Full+GDN), which interleaves GDN with a small fraction of full attention layers to maximize quality, surpassing the standard looped transformer in both performance and efficiency. We also show how to convert a pre-trained LT into an LT2-hybrid model. With about 1B tokens of training, our converted model, Ouro-hybrid-1.4B, outperforms industry-level 1B models and is competitive with industry-level 4B models while retaining the speed benefits of linear-time attention. Together, these results show a clear path toward making looped transformers more scalable and advancing efficient, capable small language models.

cs.LG↗
arXiv2026

LaDi-RL: Latent Diffusion Reasoning Prevents Entropy Collapse in Reinforcement Learning

Reinforcement learning has become a central paradigm for improving LLM reasoning, but most existing methods optimize policies over discrete token sequences. This creates a mismatch between the optimization space and the structure of reasoning: many important decisions are semantic, global, and trajectory-level rather than local token choices. Continuous latent-space RL offers a promising alternative by allowing policies to explore higher-level reasoning representations. However, simply moving to latent space is not sufficient. The resulting policy must model a complex, multi-modal distribution over valid reasoning trajectories. We therefore propose Latent Diffusion Reasoning with Reinforcement Learning (LaDi-RL), where a diffusion model generates latent reasoning trajectories through iterative denoising. This formulation enables structured exploration and expressive distribution modeling, but also introduces a fundamental credit-assignment challenge: the policy acts in latent space, while rewards are observed only after the latent is decoded into text. A naive rollout strategy therefore entangles latent reasoning quality with text decoding quality, making it unclear whether an incorrect answer results from a poor latent trajectory or from an imperfect textual realization. To address this, we introduce hierarchical latent-text rollouts. We sample multiple text completions for each latent trajectory and aggregate their rewards to obtain a decoder-marginalized estimate of latent utility. This provides a cleaner and lower-variance reward signal for optimizing the diffusion policy. Empirically, LaDi-RL outperforms token-level RL by 9.4% on code generation and 5.7% on math reasoning in pass@1, and even surpasses the base model's pass@k performance.

cs.LG↗
arXiv2026

IndustryBench: Probing the Industrial Knowledge Boundaries of LLMs

In industrial procurement, an LLM answer is useful only if it survives a standards check: recommended material must match operating condition, every parameter must respect a regulated threshold, and no procedure may contradict a safety clause. Partial correctness can mask safety-critical contradictions that aggregate LLM benchmarks rarely capture. We introduce IndustryBench, a 2,049-item benchmark for industrial procurement QA in Chinese, grounded in Chinese national standards (GB/T) and structured industrial product records, organized by seven capability dimensions, ten industry categories, and panel-derived difficulty tiers, with item-aligned English, Russian, and Vietnamese renderings. Our construction pipeline rejects 70.3% of LLM-generated candidates at a search-based external-verification stage, calibrating how unreliable industrial QA remains after LLM-only filtering. Our evaluation decouples raw correctness, scored by a Qwen3-Max judge validated at $κ_w = 0.798$ against a domain expert, from a separate safety-violation (SV) check against source texts. Across 17 models in Chinese and an 8-model intersection over four languages, we find: (i) the best system reaches only 2.083 on the 0--3 rubric, leaving substantial headroom; (ii) Standards & Terminology is the most persistent capability weakness and survives item-aligned translation; (iii) extended reasoning lowers safety-adjusted scores for 12 of 13 models, primarily by introducing unsupported safety-critical details into longer final answers; and (iv) safety-violation rates reshuffle the leaderboard -- GPT-5.4 climbs from rank 6 to rank 3 after SV adjustment, while Kimi-k2.5-1T-A32B drops seven positions. Industrial LLM evaluation therefore requires source-grounded, safety-aware diagnosis rather than aggregate accuracy. We release IndustryBench with all prompts, scoring scripts, and dataset documentation.

cs.AI↗
arXiv2026

Primal Generation, Dual Judgment: Self-Training from Test-Time Scaling

Code generation is typically trained in the primal space of programs: a model produces a candidate solution and receives sparse execution feedback, often a single pass/fail bit. Test-time scaling enriches the inference procedure by sampling multiple candidates and judging among them, but the comparative information this process reveals is discarded after inference. We argue that this information defines a dual judgment space that provides a far richer training signal: the model learns not from an isolated success or failure, but from the relative correctness structure across its own plausible attempts, identifying which succeed, which fail, and what distinguishes them. We introduce DuST (Dual Self-Training), a framework for self-training from the dual judgment space. DuST samples candidate programs from the model's own distribution, labels them through sandbox execution, retains groups containing both successes and failures, and trains the model to rank candidates by execution correctness using GRPO. The objective is purely discriminative: the model is never directly rewarded for generating correct programs. Dual self-training improves both judgment and generation. Across five models spanning two families and three scales (4B to 30B), DuST consistently improves Best-of-4 test-time scaling on LiveCodeBench. For Qwen3-30B-Thinking on LiveCodeBench v6, judgment quality improves by +6.2 NDCG, single-sample pass@1 improves by +3.1, and Best-of-4 accuracy improves by +4.1. The trained model's single rollout matches the base model's Best-of-4 performance. SFT on the same ranking data improves judgment without improving generation, confirming that on-policy RL is the mechanism that transfers dual-space learning back into primal generation.

cs.LG↗