Ziyun Zhang
Publications
Envs-FORGE: Frontier-Optimized Reward-Grounded Environment Synthesis for Agent RL
Reinforcement learning (RL) for terminal agents needs executable training environments with reliable rewards and useful difficulty. Fixed recipes such as few-shot, Self-Instruct, and Evol-Instruct apply the same prompting policy to every seed, even when the current policy would benefit from a harder, easier, or simply different task. We present Envs-FORGE, a prompting policy that converts verifier rewards into per-seed environment-synthesis actions. Envs-FORGE estimates seed pass rates, scores six projection--direction actions around a target learning frontier, and solves a per-seed mixed-integer linear program (MILP) to choose the action that conditions generation. The selected action drives synchronized rewriting of the instruction, fixtures, oracle solution, tests, and Docker environment; only gold-verified bundles enter RL training. The indexed MILP form also supports optional soft skill coverage for portfolio planning. On Qwen 3.5 35B, Envs-FORGE improves Pass@1 over Base by 9.2 percentage points on tb-core (40.0% to 49.2%) and 6.4 points on tb-2.0 (23.0% to 29.4%), exceeding the strongest fixed-recipe baseline by 2.4 and 2.1 points. It reaches 77.1% on SWE-bench Verified versus 73.4% for Base, and improves tb-core by 6.8--9.2 points across the evaluated 4B--35B models. All synthesis methods export 100 verified environments and use 2.27M--2.88M synthesis tokens, placing the comparison at the same downstream training-set size and the same operational scale. The source code is available at https://github.com/DataArcTech/DataArc-SynData-Toolkit/.
Intern-S2-Mobius: Foundation Model with Decoupled Knowledge and Reasoning
We introduce Mobius-v0, an architecture that comprises a globally shared Memory (FFN) that stores knowledge vectors and multiple Reasoners (Self-Attn) that iteratively achieve compositional reasoning. Using hidden states as cache and carrier, reasoners repeatedly query memory for required knowledge-vectors, while the knowledge is transmitted back to reasoning operators. Through this knowledge-reasoning-separation architecture, Mobius achieves better knowledge compression and reasoning efficiency. Built upon Mobius-v0 architecture: 1) Our 7B model trained-from-scratch achieves similar downstream score as a 7B Transformer baseline with 62.6% of baseline's training data. 2) Our Intern-S2-Mobius, continually-pretrained from Qwen3.5-35B, achieves similar downstream score while delivering nearly 4x end-to-end inference speedup.
On the Robustness of Temporal Vision-Language Models for Surgical Endoscopy Videos
Temporal vision-language models (TVLMs) offer a reusable, prompt-based interface for surgical video understanding, yet, their robustness under clinically realistic acquisition artifacts in endoscopy remains insufficiently characterized. In practice, degradations such as defocus, haze, motion blur, noise, cautery smoke, and packet loss introduce structured distribution shifts which may compromise video-text alignment. We study the robustness of temporal VLMs under such shifts caused by corruptions in clip frames. We introduce Endo-C6, a compact corruption benchmark of six endoscopy-realistic perturbations evaluated at a fixed high severity, and apply it to public Gastrointestinal (GI) endoscopy and laparoscopic cholecystectomy videos. Under a standardized prompt protocol, we benchmark 3 recent surgical TVLM baselines and analyze robustness in both mean and worst-case settings, spanning 294 dataset-level evaluations. Finally, we present RobustEndoCLIP, obtained by few-shot parameter-efficient tuning with VeRA, outperforming existing TVLM baselines. Our findings show that off-the-shelf TVLMs can exhibit severe worst-case collapse under endoscopy-specific corruptions, whereas lightweight few-shot adaptation can substantially improve corrupted performance and robustness without changing the prompt-based interface. We expect Endo-C6 to support standardized robustness reporting and promote more reliable clinical vision-language systems.
Meteorology-driven Causal Nowcasting of Fugitive Landfill Emissions Enables Proactive Public Health Response
Fugitive emissions from waste sites increasingly expose communities to toxic and odorous gases, yet public-health responses remain largely retrospective, with episodes investigated only after residents have been exposed. Here we show that the meteorological drivers of elevated hydrogen sulphide (HS) at a long-monitored European landfill, and the timescales over which they act, can be identified directly from routine monitoring data. We introduce CAIRN (Causal-Anchored Inference for Receptor Nowcasting), a machine-learning framework whose internal memory is matched to these measured timescales: a fast component tracking hour-scale wind-borne transport and a slow component tracking multi-hour weather changes. Trained to predict gas measurements, CAIRN operates using only routine weather variables and the calendar, without hand-engineered features. Its behaviour is consistent with the identified transport mechanisms, and the framework transfers unchanged to a second monitoring station and to co-emitted methane. Combining four such nowcasters produces a site-level, tiered alert aligned with WHO odour guidance that closely reproduces the alert generated by a direct sensor network and tracks an independent record of community odour complaints. Weather-driven nowcasting can therefore estimate community impact as an emission episode unfolds, providing public-health authorities with a validated, graded trigger for intervention and enabling exposure to be reduced during events rather than after them.
Zero-Shot Skeleton-Based Action Anticipation
Action anticipation (AA) aims to recognize ongoing human or humanoids actions from partial observations, enabling robots to predict intentions before the actions are completed. Although skeleton-based AA offers efficiency advantages, existing approaches assume that all action classes are seen during training, which limits their deployment in real-world scenarios where novel actions inevitably arise. To address this gap, we study the new task of Zero-Shot Skeleton-Based Action Anticipation (ZS-SkAA). This task requires recognizing unseen action classes using only limited early-stage skeleton sequences, combining the challenges of partial observations, temporal dynamics, and zero-shot generalization. To establish foundational research for ZS-SkAA, we introduce:(1) A baseline model comprising a spatio-temporal feature extractor and a mutual information estimation and maximization module. This baseline model explicitly aligns partial visual features with semantic class embeddings across modalities by estimating and maximizing their mutual information, enhancing generalization to unseen classes.(2) A benchmark protocol using the NTU RGB+D dataset, which is adapted for rigorous ZS-SkAA evaluation. Experiments demonstrate the effectiveness of our model as a strong baseline for ZS-SkAA, achieving high zero-shot accuracy on NTU RGB+D. This work establishes ZS-SkAA as a vital research direction for real-world systems requiring generalization to novel actions.
MathForm: Scaling Mathematical Autoformalization with Knowledge Retrieval and Verification-Guided Refinement
Autoformalization is commonly framed as translating natural-language mathematical statements into machine-verifiable formal languages such as Lean 4. However, faithful formalization requires more than translation. Models must map mathematical concepts to the complex hierarchy of types and definitions in formal libraries such as Mathlib, while ensuring that generated statements preserve the meaning of the source propositions. Existing approaches struggle because they rely heavily on the model's parametric memory for library-specific knowledge, while common data construction pipelines often resort to filtering single-pass outputs and lack mechanisms for feedback-driven revision. To address these challenges, we introduce MathForm, an autoformalization framework for constructing verified training data through Mathlib knowledge retrieval and verification-guided iterative refinement. Before generation, a retrieval planner gathers relevant definitions and existing formalizations from Mathlib to guide the formalization generator. Generated statements are then revised using compiler diagnostics and semantic-consistency feedback. Using this framework, we construct FormalVerse, a Lean 4 dataset containing approximately 367K verified examples across diverse mathematical domains and sources. We then train MathForm-8B through supervised fine-tuning followed by reinforcement learning. Across six benchmarks, MathForm-8B achieves average Pass@8 rates of 88.06% under Syntax Check (SC) and 72.37% under Consistency Check (CC), outperforming multiple specialized 32B autoformalizers. On the challenging FATE-H and FATE-X subsets, it attains CC pass rates of 63% and 37%, exceeding the strongest specialized baselines in both cases.
APTER: Adaptive Post-Training with Expert-Grounded Rubrics
As large language models enter professional domains, they must satisfy domain constraints, include critical evidence, and provide complete reasoning rather than merely produce fluent responses. Existing post-training methods often rely on holistic preferences or outcome-level verification, while recent rubric-based methods usually generate rubrics independently for each query. In specialized domains, such unconstrained rubrics may omit critical requirements and vary across samples, hindering the diagnosis and targeted repair of persistent capability deficiencies. We propose APTER (Adaptive Post-Training with Expert-Grounded Rubrics), a framework that integrates structured domain knowledge into fine-grained evaluation, optimization, and diagnosis for specialized complex reasoning. First, expert-grounded rubric construction starts from an expert criteria framework built by domain experts, where each criterion represents a stable professional capability. For each query, APTER selects relevant criteria and instantiates them into query-level rubrics linked to their source criteria, turning reusable expert criteria into executable query-level supervision without reference answers. Second, adaptive post-training uses rubric verdicts as both optimization and criterion-level diagnostic signals. Aggregating low-scoring verdicts by criterion ID reveals persistent deficiencies and triggers targeted supervised fine-tuning updates during reinforcement learning. Experiments on mathematical reasoning and medical question answering show consistent gains across both domains. Across three model generations, APTER improves the mathematics and medical averages over the corresponding base models by up to 15.86 and 8.04 points, respectively. Code and rubric datasets are available at https://github.com/AntDT-APTER/APTER.
How Much Do Legal RAG Systems Still Hallucinate?
Hallucination is a major challenge for retrieval-augmented generation (RAG) systems in the legal domain, where ungrounded answers can lead to serious consequences. To better understand this problem, we conduct a fine-grained analysis of hallucination behavior in eight legal RAG systems across two legal corpora, the GDPR (in English) and a national civil law (in French). Using claim-level and answer-level evaluation, we report on hallucination density and severity, analyze performance across question categories and user personas, and validate our findings on an independent set of 142 legal-expert-authored questions. Our results show that hallucinations remain pervasive, ranging from less than 10% of responses for the best-performing systems to nearly half in the worst case. We further find that false-premise questions, containing incorrect assumptions that must be rejected, produce high hallucination rates on the manually-drafted questions.
MMUSV-Sim: A Perception-Oriented Simulation and Data-Generation Platform for Multi-USV Cooperative Perception
Cooperative perception among multiple unmanned surface vehicles (USVs) combines complementary observations to extend maritime target sensing beyond the view range and field of a single platform. Developing such systems at scale calls for a unified workflow for configurable multi-USV scenarios, multimodal acquisition, and shared annotations. We present MMUSV-Sim, a perception-oriented maritime simulation and data-generation platform built on Unreal Engine 5 and Project AirSim. It provides island, open-sea, and port environments; configurable weather, time of day, and wave conditions; a diverse vessel asset library; and spline-based multi-vessel motion. MMUSV-Sim acquires RGB, depth, semantic, LiDAR, and radar observations across multiple USVs and captures a common world state for per-agent annotation export. Experiments verify that the configured wave settings produce the intended changes in vessel heave, roll, and pitch, and evaluate the geometric consistency between projected annotations and semantic renderings. In LiDAR-based cooperative BEV vessel detection experiments on the generated multi-USV dataset, Early Fusion achieves an AP@0.5 of 72.74, compared with 45.54 using a single USV.
OccPlanner: Goal-Aware Occupancy-Conditioned Diffusion Planner for Pixel-Goal Navigation
Pixel-goal navigation specifies targets directly in the agent's camera view, but a target pixel provides neither metric depth nor traversability, making 3D goal grounding and collision-free continuous planning challenging. We present OccPlanner, a goal-aware occupancy-conditioned diffusion planner that grounds pixel goals in egocentric metric space and sequentially conditions the goal representation on temporal visual context and learned local 3D occupancy features. To provide occupancy supervision at scale, we introduce L3ROcc, which converts monocular RGB navigation videos into robot-centric local 3D occupancy annotations through geometric reconstruction and ray-based visibility reasoning. We train OccPlanner on InternData-N1 and evaluate it in closed-loop simulation across four unseen scene categories from InternScenes and two goal-distance ranges. In the 5-8 m setting, OccPlanner increases the average success rate (SR) over NavDP from 20.81% to 71.55% across the four categories, reaching 86.20% and 84.92% in cluttered-easy and cluttered-hard scenes, respectively. Real-world open-loop experiments on a Unitree Go2 further provide initial evidence of sim-to-real transfer and adaptation with L3ROcc-generated supervision.