Meituan & ICML'26 Paper Sessions: Generalist Agents and Video Generation

MemOCR: Layout‑Aware Visual Memory for Efficient Long‑Horizon Reasoning

PDF: https://arxiv.org/abs/2601.21468

Long‑horizon agent reasoning requires compressing ever‑growing interaction histories into a limited context window. Conventional memory systems serialize history as text, incurring a uniform token‑level cost that grows linearly with length. MemOCR proposes a multimodal memory agent that allocates memory space adaptively via visual layout, increasing information density under tight context budgets. Experiments on long‑context multi‑hop and single‑hop QA benchmarks demonstrate that MemOCR surpasses strong text baselines, particularly when the context budget is extremely limited.

ScaleEnv: Scaling Environment Synthesis from Scratch for Generalist Interactive Tool‑Use Agent Training

PDF: https://arxiv.org/abs/2602.06820

Training generalist agents that can adapt to diverse scenarios demands fully programmable interactive environments and verifiable tasks. ScaleEnv builds such an environment and task suite from the ground up. Programmatic tests guarantee environment reliability, while tool‑dependency graphs and executable‑action verification ensure task completeness and solvability. Evaluation on unseen multi‑round tool‑use benchmarks shows significant performance gains, highlighting strong generalisation.

V_0: A Generalist Value Model for Any Policy at State Zero

PDF: https://arxiv.org/abs/2602.03584

Value models for LLM‑based reinforcement learning face a coupling dilemma: they must be synchronised with constantly updating policies. V_0 decouples value estimation from specific policy parameters by redefining the task as a context‑learning problem that predicts the performance of unseen policies. Experiments reveal that V_0 tracks policy evolution better than coupled value models during GRPO training, optimises cold‑start budget allocation, and approaches the performance‑cost Pareto frontier in inference routing.

Learning to Self‑Verify Makes Language Models Better Reasoners

PDF: https://arxiv.org/abs/2602.07594

Large language models generate promising reasoning paths for complex tasks but remain weak at verifying their own answers. A multi‑task reinforcement‑learning framework jointly optimises generation and self‑verification as complementary objectives. Empirical results show that this approach outperforms generation‑only training on both generation quality and verification capability.

AgentNoiseBench: Benchmarking Robustness of Tool‑Using LLM Agents Under Noisy Conditions

PDF: https://arxiv.org/pdf/2602.11348

Existing benchmarks do not capture robustness of tool‑using agents under imperfect user instructions and unreliable tool feedback. AgentNoiseBench models two primary noise sources—user‑side instruction noise and tool‑side result noise—provides a modular noise‑injection pipeline and multidimensional evaluation metrics. Evaluation of 25 tool‑using models indicates that tool‑side noise typically causes larger performance drops than user‑side noise.

AJ‑Bench: Benchmarking Agent‑as‑a‑Judge for Environment‑Aware Evaluation

PDF: https://arxiv.org/abs/2604.18240

Scaling RL‑driven LLM agents raises challenges for reliable behaviour verification in complex environments. Rule‑based validators and LLM‑as‑Judge models struggle to generalise beyond narrow domains. AJ‑Bench introduces a benchmark that evaluates agents acting as judges by actively interacting with environments and tools to gather verifiable evidence. The benchmark covers search, data‑system, and GUI domains (155 tasks, 516 annotated trajectories). Experiments show that agent‑as‑judge methods achieve stable gains over LLM‑as‑Judge baselines while exposing open challenges.

LUVE: Latent‑Cascaded Ultra‑High‑Resolution Video Generation with Dual Frequency Experts

PDF: https://arxiv.org/abs/2602.11564

To reconcile ultra‑high‑resolution video generation with coherence and compute constraints, LUVE proposes a three‑stage latent‑cascaded framework using dual‑frequency experts. Stage 1 generates low‑resolution video to preserve motion consistency; Stage 2 upsamples in latent space, drastically reducing memory and compute; Stage 3 fuses high‑ and low‑frequency experts to refine semantics and details. Experiments demonstrate superior realism and fidelity, and the core ideas have been applied to the LongCat‑Video model.

Infinite‑World: Scaling Interactive World Models to 1000‑Frame Horizons via Pose‑Free Hierarchical Memory

PDF: https://arxiv.org/abs/2602.02393

Infinite‑World targets long‑horizon interactive world modelling in real‑world videos. It introduces a pose‑free hierarchical memory compressor that stores history latents within a fixed budget, reducing long‑range modelling cost. Uncertainty‑aware action annotation improves learning under noisy trajectories, and high‑revisit data fine‑tuning enhances loop‑closure. The result is a world model better suited for long‑term spatio‑temporal consistency.

WildActor: Unconstrained Identity‑Preserving Video Generation

PDF: https://arxiv.org/pdf/2603.00586

WildActor addresses identity drift, full‑body inconsistency, and pose artefacts in dynamic, long‑shot video generation. It builds a 1.6 M video and 18 M multi‑view image dataset (Actor‑18M) to mitigate frontal‑face bias, introduces Asymmetric Identity‑Preserving Attention (AIPA) to decouple identity from motion, and employs Identity‑aware 3D Rotational Positional Encoding (I‑ROPE) to separate spatio‑temporal tokens. Monte‑Carlo view‑adaptive sampling enables robust arbitrary‑view control. Experiments on the newly constructed Actor‑Bench show substantial gains in whole‑body consistency and text‑alignment over existing open‑source and commercial models.

Spectrum‑Adaptive Fine‑Tuning for LLMs (SAFT)

PDF: https://github.com/sjtu-scx/SAFT/blob/main/SAFT.pdf

Standard supervised fine‑tuning (SFT) optimises cross‑entropy, which approximates accuracy but can over‑fit noisy samples and become over‑confident. Direct fine‑tuning (DFT) optimises a smooth approximation of accuracy, improving robustness but reducing learning efficiency on hard samples. SAFT proposes a lightweight pre‑test protocol: train SFT and DFT on a small subset, compare validation performance, and select geometric interpolation (Geo‑SAFT) for high‑SNR data or harmonic interpolation (Har‑SAFT) for low‑SNR data. This data‑adaptive interpolation yields a better robustness‑efficiency trade‑off than linear interpolation.

TRIP‑Bench: A Benchmark for Long‑Horizon Interactive Agents in Real‑World Scenarios

PDF: https://arxiv.org/pdf/2602.01675

TRIP‑Bench constructs a travel‑planning benchmark from real‑world data, featuring 18 tools and over 40 travel constraints. It evaluates agents on maintaining global constraints, tool usage, handling user requirement changes, and iterative plan revisions across up to 15 dialogue rounds, 150+ tool calls, and >200 k tokens. Existing state‑of‑the‑art models perform limitedly; the authors introduce GTPO, a multi‑round RL method that improves Qwen2.5‑32B‑Instruct’s performance beyond Gemini‑3‑Pro.

InfVSR: Toward Consistency‑Driven Streaming Generative Video Super‑Resolution

PDF: https://arxiv.org/pdf/2510.00948

InfVSR tackles inefficiencies and temporal inconsistency of diffusion‑based video super‑resolution on long sequences. It converts a pretrained video DiT into a causal streaming architecture with a rolling KV cache for smooth local transitions, and injects global semantic anchors via cross‑attention to curb drift. Training combines block‑wise pixel supervision with cross‑block distribution matching, and distils the diffusion process into a single‑step inference. Results achieve state‑of‑the‑art performance, 58× faster inference, and constant memory usage on long sequences.

DRIVE: Distributional and Retrieval‑Augmented Bidding with Value Evaluation

PDF: https://arxiv.org/abs/2606.14192

Standard Decision Transformers struggle with average‑action traps, long‑tail hallucinations, and lack of reasoning optimisation in complex bidding environments. DRIVE proposes a generate‑retrieve‑evaluate loop: (1) replace deterministic outputs with a Gaussian mixture model to avoid policy collapse; (2) add a retrieval mechanism to enhance long‑tail memory and prevent hallucinations; (3) employ an IQL critic for real‑time evaluation of generated and historical actions. The framework markedly improves decision robustness.