arXiv subjects

Wenhao Wu

Publications and source records attributed to Wenhao Wu.

At least 19 recordsLinked to original sources

arXiv2026

Second-Best Gains from Trade in Matching Markets

We study gains from trade (GFT) in two-sided matching markets with independent private types and arbitrary downward-closed feasibility constraints. The second-best benchmark is the maximum expected GFT achievable by a Bayesian incentive-compatible, interim individually rational mechanism that is strongly budget balanced at every report profile. These constraints generally preclude attaining the first-best GFT and raise the question of how much efficiency must be lost. We prove that the second-best GFT is at least one half of the first-best GFT in every such matching market. This recovers and generalizes the recent $1/2$ guarantee for bilateral trade by Liu, Qin, Ren, and Wang to markets with multiple buyers and sellers and arbitrary downward-closed feasibility constraints. Together with their matching upper bound for bilateral trade, our result establishes a tight worst-case ratio of $1/2$ for this general class of matching markets. Our proof builds on the virtual-GFT framework of Brüstle, Cai, Wu, and Zhao to reduce the problem to a one-parameter Lagrangian. The main step is a geometric, edge-by-edge analysis based on first-best edge-selection regions, combined with a randomized contraction in rank space. The same analysis also determines the exact worst-case efficiency for every level of outside subsidy. In particular, even a small outside subsidy can significantly improve the worst-case guarantee.

cs.GT↗

arXiv2026

The Optimizer's Windfall: Bounding Hindsight Gain under Random Restrictions

Suppose an optimization problem is randomly restricted, independently on each variable, and re-optimized after the restriction is revealed. For multilinear objectives, every fixed solution is unbiased after rescaling, but re-optimizing introduces an upward bias. We bound this excess, the hindsight gain, over products of arbitrary compact local domains in Euclidean unit balls. For interactions $A=(A_e)_e$ among $r$ distinct variables and independent random weights $ζ_i$, the gain is at most $O_r(\|A\|_F M^{(r-1)/2}σ_{\max})$, where $\|A\|_F^2$ is the sum of squared coefficients, $M=\sum_i\mathbb{E}ζ_i^2$, and $σ_{\max}^2=\max_i\mathrm{Var}(ζ_i)$; only second moments are needed. Under uniform retention with probability $p$, we sharpen this to $O_r(p^{r-1/2}\sqrt{1-p}\,n^{(r-1)/2}\|A\|_F)$, which is tight for every fixed $r$ and $p\in(0,1)$. Our proof combines a comparison principle along a single support direction with tensor projection, vector contraction, and repeated thinning. Applications: (i) For fixed arity and alphabet size, solving the subproblem induced by $O(\varepsilon^{-2})$ sampled variables estimates a dense Max-$r$CSP optimum within $\pm\varepsilon n^r$, and this sample size is optimal. (ii) The same approach estimates MaxCut and the distance to bipartiteness within $\pm\varepsilon n^2$ using $O(\varepsilon^{-4})$ nonadaptive queries and $O(\varepsilon^{-2})$ examined vertices, both optimal. (iii) We answer a question of Yufei Zhao: vertex percolation at rate $p$ preserves the max-cut fraction whenever $p\,e(G)/v(G)\to\infty$, and this condition is sharp. (iv) For bounded $r$-local quantum Hamiltonians of fixed local dimension, $O(\varepsilon^{-3})$ uniformly sampled particles suffice to estimate the ground-state energy within $\pm\varepsilon n^r$, improving the previous $O(\varepsilon^{-6}\log(1/\varepsilon))$ bound.

cs.DS↗

arXiv2026

HarnessPAI: An Evolving Harness for Physical AI

Physical AI aims to build embodied agents that perceive the world, understand and reason about it, and decide how to act. Yet the field has focused primarily on the last component: the action model that maps observations to low-level controls. The prevailing training recipe can erode the perceptual and reasoning capabilities needed for robust behavior, leaving even strong action models vulnerable to scene perturbations and long-horizon tasks. We introduce HarnessPAI, a model- and embodiment-agnostic Harness framework for Physical AI that treats code as the executable and evolvable interface that organizes the underlying action primitive. The framework separates two timescales: within a rollout, it executes open-loop at the program level, with a fixed program guiding and checking execution; across rollouts, it evolves closed-loop, using execution feedback to revise the program and distill failures into reusable skills. Across desktop robot arms, household robots, a robot vacuum, and a legged walking agent, HarnessPAI demonstrates consistent improvements against baselines. Beyond the harness, the converged program also acts as a low-cost, reliable source of expert data: it generates training data for fine-tuning VLA and WAM models on tasks where it is hard for these models to obtain successful trajectories on their own. Our results suggest that the frontier of Physical AI depends not only on stronger action models, but also on executable harnesses that integrate perception, task understanding and reasoning, and action execution into a unified, verifiable, and feedback-driven system. Website: https://darwin-agent.github.io/HarnessPAI

cs.RO↗

arXiv2026

AdaStep: Adaptive Step Credit Weighting for Agentic Reinforcement Learning

Long-horizon LLM agents are typically trained with sparse outcome rewards, making trajectory-level objectives too coarse to distinguish the contribution of individual decisions. Step-level credit assignment provides finer-grained supervision, but its estimates can be unreliable because observed returns also depend on subsequent actions, environment transitions, and trajectory length. We propose AdaStep, an Adaptive Step-credit weighting method that controls how strongly each group-derived local advantage modifies the trajectory-level signal. We formulate this weighting as a mean-squared-error estimation problem for the latent step advantage and, under an explicit conditional sampling assumption, derive an optimal per-state shrinkage coefficient. The coefficient admits a signal-to-total-variance interpretation: it preserves local credit when return variation is attributable to the selected action and suppresses it when variation is dominated by downstream randomness. AdaStep requires only lightweight scalar computation, with no critic, additional rollouts, or extra model inference. Experiments with three model backbones on ALFWorld, WebShop, and ScienceWorld show consistent improvements over baselines at low computational cost.

cs.LG↗

arXiv2026

Look Before You Leap: Distilling Tree Search into Action Evaluation for Frozen VLA Models

Vision-Language-Action (VLA) models acquire broad embodied capabilities through large-scale pretraining, yet their generalization remains far more fragile than that of LLMs and VLMs. The prevailing remedy, post-training via supervised fine-tuning or reinforcement learning, improves task-specific performance but narrows the generalist capability that makes pretraining valuable. We identify a key bottleneck: VLA failures stem not only from action generation but also from action evaluation. A diagnostic pass@k study confirms that frozen VLAs already contain competent behaviors in their output distribution, with overall success rates rising from 33% at pass@1 to 92% at pass@32. Inspired by this, we propose SVA (Search, Value, and Act), a simple framework that equips frozen VLA policies with long-term consequence awareness. SVA first uses Monte-Carlo tree search in simulation to sample and explore the VLA's output distribution under a finite search budget, collecting diverse trajectories annotated with empirical returns; this knowledge is then distilled into a lightweight Q-value model that predicts the expected consequence of candidate actions; at deployment, the frozen VLA proposes multiple candidates and the evaluator selects the one with the highest uncertainty-regularized Q-value, requiring no simulator access. By decoupling action proposal from consequence evaluation, SVA preserves the generalization capacity of the VLA backbone while improving task success. Across embodied benchmarks, SVA consistently improves generalization on unseen tasks and exhibits strong test-time scaling; across real-robot manipulation tasks, it improves average success from 42.8% to 55.6% (+12.8 points). Strikingly, SVA enables a 9B VLA to outperform a 27B VLA by 7 points at 27% lower inference latency, suggesting that scaling test-time evaluation is more cost-effective than scaling model size.

cs.RO↗

arXiv2026

Beating One Half for Online Bipartite Matching with Reusable Resources

We study online bipartite matching with unit-inventory reusable resources, where requests arrive in an adversarially fixed order, and each use of a resource makes it unavailable for an independent duration drawn from a resource-dependent distribution. The benchmark knows all requests in advance but cannot observe a duration before choosing the corresponding use. The classical Ranking algorithm of Karp, Vazirani, and Vazirani (STOC 1990) fixes a uniformly random priority order of the resources and matches each arriving request to its highest-priority available neighbor. It achieves the optimal competitive ratio $1-1/e$ for unweighted nonreusable resources, but whether it beats $1/2$ for reusable resources has remained open. We prove that, for unweighted resources with resource-dependent stochastic durations, Ranking achieves a competitive ratio of $(5-2\sqrt3)/3\approx0.511966$. We also give a black-box reduction from unweighted Ranking to resource-weighted matching: any unweighted competitive ratio $α>1/2$ yields a weighted ratio strictly above $1/2$. With independent sampling access to the duration distributions, the reduction gives a weighted ratio of $0.500034$. These results resolve two questions left open by Delong et al. (MOR 2024): whether Ranking beats $1/2$, and whether one can beat $1/2$ under stochastic durations. We analyze Ranking resource by resource, rather than request by request. For deterministic durations, this gives a reduction to random-order greedy for a coverage function. We then extend the analysis to stochastic durations by comparing the residual schedules of Ranking and a greedy algorithm, and apply a finer analysis of the random ranks to obtain the stated $0.511$ bound. For the weighted reduction, we apply Ranking within groups of similar weights and uses weighted greedy to control the loss between groups.

cs.DS↗

arXiv2026

Text-to-3D Policy: Fine-Grained Language-Behavior Alignment for Unseen Specification Generalization

3D visuomotor policies provide a strong foundation for spatially precise manipulation, yet current text-to-3D policies struggle to follow unseen fine-grained behavioral specifications beyond those covered by demonstrations. We study this challenge as unseen specification generalization, where language specifies behaviorally significant variations, such as target position, displacement, or articulated state, that are absent from policy training. We find that pretrained language representations and conventional global behavior-language alignment capture coarse task semantics but often blur nearby specifications that require distinct behaviors. We introduce T3DP, a Text-to-3D Policy framework for fine-grained language-behavior alignment. Rather than compressing each instruction and demonstration into a single global embedding, T3DP preserves their local structures and establishes bidirectional token-level correspondence between linguistic elements and behavioral segments. This directly grounds subtle linguistic variations in the behavior components they affect, preventing closely related specifications from collapsing in the representation space. The resulting specification-sensitive language representation conditions a point-cloud-based 3D diffusion policy, enabling more precise control over unseen behavioral specifications without modifying the underlying policy architecture. Across Meta-World, ManiSkill, and RoboTwin, T3DP improves average held-out-specification success over global language-behavior alignment by +11.0-14.2 points, with gains on all 15 task families; on real-robot tasks, it further raises average success from 47.5% to 65.0% (+17.5 points). Representation and action-probe analyses show that fine-grained alignment better preserves specification geometry and action-relevant variation, linking local behavior grounding to downstream control.

cs.RO↗

arXiv2026

The Power of Simple Mechanisms: A Tight $e$-Approximation for Gains from Trade in Matching Markets

We study how well simple mechanisms approximate gains from trade (GFT) in two-sided matching markets with independent buyer values and seller costs, where feasible outcomes form an arbitrary downward-closed family of matchings. This model includes bilateral trade and double auctions as special cases. We focus on the Generalized Random-Offerer (GRO) mechanism, which is an equal mixture of the Generalized Sellers-Offering Mechanism (GSOM) and the Generalized Buyers-Offering Mechanism (GBOM). We determine GRO's exact worst-case approximation ratio with respect to first-best GFT, showing that it is $e$. This improves the previous $3.15$ approximation guarantee (Babaioff et al. STOC 2026) for the same mechanism. In bilateral trade, the result implies a $1/e$ guarantee for the random-offerer mechanism, improving the previous $1/π$ bound (Jo 2026). Our analysis uses a two-dimensional inequality in quantile space to compare first-best GFT with optimal one-sided auction profits. For each potential trade, we identify the region of buyer values and seller costs for which the first-best allocation selects that trade, and then randomly shrink this region in quantile space, yielding posted-price rules whose expected profits can be evaluated exactly. We establish tightness of the $e$ approximation by constructing markets with regular type distributions, pairwise disjoint trading edges, and a single knapsack constraint. On these instances, GRO's expected GFT approaches a $1/e$ fraction of first-best GFT, therefore ruling out any better guarantee even under these restrictions.

cs.GT↗

arXiv2026

MinCU: A Fine-Grained Benchmark for Grounded Minimal-Change Understanding in Image Pairs

Localizing and describing fine-grained differences between near-identical images is a critical yet underexplored capability for multimodal large language models (MLLMs). Existing benchmarks largely assess semantic comparison or single-image grounding in isolation, without jointly requiring faithful description and physical localization. To bridge this gap, we introduce MinCU, a benchmark for grounded minimal-change understanding, where each sample consists of an image pair differing by a single atomic variation in object category, attribute, count, or spatial position, and models are evaluated on their ability to describe the change, localize the changed regions, and identify the changed entity. We further propose Semantic-Guided Implicit Spatial Anchors (SG-ISA), a structured autoregressive method that decomposes prediction into a Think-Locate-Describe sequence. SG-ISA first predicts a semantic cue for the changed concept, then uses discrete spatial anchors as an implicit localization scaffold, and finally generates the change description together with the grounding box. Experiments reveal that even the strongest closed-source MLLMs and recent R1-style reasoning models struggle on MinCU, with most failing to jointly produce accurate descriptions and grounding boxes. Compared to the previous chain-of-thought method, fine-tuning with SG-ISA yields substantial joint improvements in grounding accuracy and description quality while reducing reasoning-token overhead by approximately 26%. These results suggest that an implicit intermediate spatial interface can be more effective than relying solely on model scale for grounded dual-image understanding.

cs.CV↗

arXiv2026

TRACE: A Self-Evolving Skill Bank for Consistent, Limit-Aware LLM Agents

Reliable deployment of LLM agents in user-facing products depends not on raw task-solving ability but on consistency and limit-awareness: behaving the same way across repeated trials, and recognizing when a request cannot, or cannot yet, be safely fulfilled. CAR-bench exposes this reliability gap in the domain of in-car assistants: an LLM-simulated user issues incomplete or ambiguous requests, requiring the agent to resolve uncertainty through multi-turn dialogue and tool use while strictly adhering to domain policies. Even frontier models show a substantial gap between what they can solve at least once (Pass@3) and what they solve consistently across trials (Pass^k). We bridge this gap with TRACE (TRAjectory-Contrastive Evolution), which iteratively improves a skill-based agent's behavioral knowledge without modifying model weights. This knowledge is organized as a Skill Bank of modular, retrievable skills, each encoding a self-contained set of tool-use rules and behavioral guidelines. TRACE evolves this bank through an agentic self-evolution loop: after each evaluation round, it groups trajectories by the skills invoked and refines each skill by contrasting successful and failed behaviors. The updated bank then guides subsequent rounds, while during deployment the Actor performs state-conditioned skill orchestration at every turn. On GPT-5.5, TRACE improves consistency (Pass^3) by 34.6 points, from 59.9% to 94.5%, while shrinking the gap between potential and reliable performance to just 4.0 points. On the official hidden set, TRACE achieved first place using GPT-5.6-Sol, attaining a Pass^3 score of 70%-a 40% relative improvement over the baseline. These results show that TRACE converts high model potential into stable, consistent performance gain. Project homepage: https://darwin-agent.github.io/Car-bench-TRACE.

cs.CL↗

arXiv2026

LowRankArena: A Standardized Evaluation Platform for SVD-Based LLM Compression

SVD-based low-rank compression has become a fast-growing direction for reducing the memory and computational cost of large language models (LLMs). However, meaningful comparison across existing studies remains difficult as prior evaluations use varied benchmarks, inconsistent ratios, and diverse setups, often failing to isolate low-rank effects from auxiliary techniques. As a result, it remains unclear whether reported gains reflect method-level improvements or differences in evaluation protocol. This lack of comparability highlights the need for a unified, reproducible evaluation platform. To address this problem, we present LowRankArena, a standardized evaluation platform for SVD-based LLM compression. LowRankArena unifies task versions, uniform-precision compression budgets, comparison regimes, and inference measurements, and provides a reproducible pipeline with over 3 TiB released compressed checkpoints. Using LowRankArena, our aligned audit of five representative SVD methods reveals that prior findings are highly conditional under standardized protocols: clear leaders and performance tiers shift across backbones and keep ratios, multiple-choice accuracy can hide large perplexity degradation, and nominal low-rank savings yield workload-dependent and often limited end-to-end speedups. Our code is available at: https://github.com/Zishan-Shao/lowrankarena.git.

cs.CL↗

arXiv2026

MemFuse: Multi-Source Memory Fusion from Fragmented Observations

Long-term memory is essential for agents that operate across extended interactions, yet existing memory systems and benchmarks predominantly focus on single-source textual histories. In realistic settings, however, relevant information is often fragmented across applications and devices, as well as across users and time, requiring agents to integrate dispersed observations into coherent episodic memories while preserving their source provenance. To address these gaps, we introduce **MemFuseBench**, a benchmark for *multi-source memory fusion*. MemFuseBench is built with a Scene-to-Sensor pipeline that synthesizes controllable scenarios into source-tagged observations, evidence-grounded questions, and adversarial distractors. It enables systematic evaluation of temporal reasoning, cross-source evidence fusion, and robustness to noise. We further propose **MemFuse**, a structured memory system that preserves source-level evidence in event-layer atomic memory and organizes related atomic events into cluster-layer fused memory within a causal fusion graph. During retrieval, MemFuse retrieves and organizes related evidence fragments while maintaining traceability to original source events. Experiments on MemFuseBench show that MemFuse achieves the best overall performance among the evaluated memory systems under all three LLM settings and consistently improves performance on questions requiring cross-source evidence fusion.

cs.CL↗

arXiv2026

ShadowProbe: Language-Extensible Detection of Hidden Algorithmic Complexity Vulnerabilities

Algorithmic Complexity Vulnerabilities (ACVs) arise when adversarial inputs trigger worst-case execution behavior, causing severe performance degradation or Denial-of-Service conditions. A key but underexplored source is shadow complexity: non-trivial computational costs hidden inside seemingly benign standard library APIs. Because these costs are invisible at call sites, attackers can exploit them to induce unexpected superlinear runtime behavior. Existing ACV detectors often rely on fuzzing, symbolic execution, or hybrid analysis, but they are usually language-specific, require substantial manual effort to construct harnesses, and depend on heavy runtime instrumentation. We present ShadowProbe, a scalable and language-extensible framework for discovering ACVs through lightweight static analysis, automated reconstruction of execution contexts, and Large Language Model (LLM) assisted test generation. ShadowProbe uses a structured multi-stage pipeline: it statically screens for candidate functions guided by shadow-complexity signals, reconstructs minimal executable contexts from project-level symbols, and synthesizes size-controlled inputs to probe worst-case behavior. It then validates candidates using execution-time measurements and robust statistical growth inference, separating true algorithmic blowups from runtime noise such as garbage collection and JIT compilation effects. We evaluate ShadowProbe on the WISE benchmark, where it consistently improves analysis efficiency over existing approaches. We further apply it to large-scale systems including CPython, the JDK, Zig, Rustc, and vLLM, uncovering many previously unknown ACVs, many of which have been confirmed and partially remediated by maintainers. These results show that ShadowProbe can identify hidden algorithmic risks across diverse real-world codebases.

cs.CR↗

arXiv2026

From Conflict to Consensus: Boosting Medical Reasoning via Multi-Round Agentic RAG

Large Language Models (LLMs) exhibit high reasoning capacity in medical question-answering, but their tendency to produce hallucinations and outdated knowledge poses critical risks in healthcare fields. While Retrieval-Augmented Generation (RAG) mitigates these issues, existing methods rely on noisy token-level signals and lack the multi-round refinement required for complex reasoning. In this paper, we propose MA-RAG (Multi-Round Agentic RAG), a framework that facilitates test-time scaling for complex medical reasoning by iteratively evolving both external evidence and internal reasoning history within an agentic refinement loop. At each round, the agent transforms semantic conflict among candidate responses into actionable queries to retrieve external evidence, while optimizing history reasoning traces to mitigate long-context degradation. MA-RAG extends the self-consistency principle by leveraging the lack of consistency as a proactive signal for multi-round agentic reasoning and retrieval, and mirrors a boosting mechanism that iteratively minimizes the residual error toward a stable, high-fidelity medical consensus. Extensive evaluations across 7 medical Q&A benchmarks show that MA-RAG consistently surpasses competitive inference-time scaling and RAG baselines, delivering substantial +6.8 points on average accuracy over the backbone model. Our code is available at https://github.com/NJU-RL/MA-RAG.

cs.CL↗

arXiv2026

FlashSVD v1.5: Making Low-Rank Transformers Inference Actually Fast

SVD-based Low-rank compression reduces transformer parameters and nominal FLOPs, but these savings often translate poorly into real LLM serving speedups. We show that this gap is largely a runtime problem: factorized checkpoints fragment execution paths, and the resulting overhead differs substantially between prefill and autoregressive decode. We present FlashSVD v1.5, a unified inference runtime for serving SVD-compressed transformers. FlashSVD v1.5 maps diverse public SVD compression families to a common factorized representation and combines phase-specific kernels with dense-KV decode, packed MLP execution, and per-layer CUDA-graph replay to reorganize the low-rank serving path into a thin runtime. Across representative decoder-serving settings, FlashSVD v1.5 achieves up to 2.55x decode and 2.39x end-to-end speedup, and it attains 1.48x average decode and 1.44x average end-to-end speedup across multiple popular SVD compression families. These results suggest that practical low-rank acceleration requires runtime co-design, not compression algorithms alone. Our code is available at: https://github.com/Zishan-Shao/FlashSVD.

cs.LG↗

arXiv2026

CODER: Coupled Diversity-Sensitive Momentum Contrastive Learning for Image-Text Retrieval

Image-Text Retrieval (ITR) is challenging in bridging visual and lingual modalities. Contrastive learning has been adopted by most prior arts. Except for limited amount of negative image-text pairs, the capability of constrastive learning is restricted by manually weighting negative pairs as well as unawareness of external knowledge. In this paper, we propose our novel Coupled Diversity-Sensitive Momentum Constrastive Learning (CODER) for improving cross-modal representation. Firstly, a novel diversity-sensitive contrastive learning (DCL) architecture is invented. We introduce dynamic dictionaries for both modalities to enlarge the scale of image-text pairs, and diversity-sensitiveness is achieved by adaptive negative pair weighting. Furthermore, two branches are designed in CODER. One learns instance-level embeddings from image/text, and it also generates pseudo online clustering labels for its input image/text based on their embeddings. Meanwhile, the other branch learns to query from commonsense knowledge graph to form concept-level descriptors for both modalities. Afterwards, both branches leverage DCL to align the cross-modal embedding spaces while an extra pseudo clustering label prediction loss is utilized to promote concept-level representation learning for the second branch. Extensive experiments conducted on two popular benchmarks, i.e. MSCOCO and Flicker30K, validate CODER remarkably outperforms the state-of-the-art approaches. Our code is available at: https://github.com/BruceW91/CODER.

cs.CV↗

arXiv2025

A Comprehensive Survey on Long Context Language Modeling

Efficient processing of long contexts has been a persistent pursuit in Natural Language Processing. With the growing number of long documents, dialogues, and other textual data, it is important to develop Long Context Language Models (LCLMs) that can process and analyze extensive inputs in an effective and efficient way. In this paper, we present a comprehensive survey on recent advances in long-context modeling for large language models. Our survey is structured around three key aspects: how to obtain effective and efficient LCLMs, how to train and deploy LCLMs efficiently, and how to evaluate and analyze LCLMs comprehensively. For the first aspect, we discuss data strategies, architectural designs, and workflow approaches oriented with long context processing. For the second aspect, we provide a detailed examination of the infrastructure required for LCLM training and inference. For the third aspect, we present evaluation paradigms for long-context comprehension and long-form generation, as well as behavioral analysis and mechanism interpretability of LCLMs. Beyond these three key aspects, we thoroughly explore the diverse application scenarios where existing LCLMs have been deployed and outline promising future development directions. This survey provides an up-to-date review of the literature on long-context LLMs, which we wish to serve as a valuable resource for both researchers and engineers. An associated GitHub repository collecting the latest papers and repos is available at: \href{https://github.com/LCLM-Horizon/A-Comprehensive-Survey-For-Long-Context-Language-Modeling}{\color[RGB]{175,36,67}{LCLM-Horizon}}.

cs.CL↗

arXiv2025

Text-to-Decision Agent: Offline Meta-Reinforcement Learning from Natural Language Supervision

Offline meta-RL usually tackles generalization by inferring task beliefs from high-quality samples or warmup explorations. The restricted form limits their generality and usability since these supervision signals are expensive and even infeasible to acquire in advance for unseen tasks. Learning directly from the raw text about decision tasks is a promising alternative to leverage a much broader source of supervision. In the paper, we propose \textbf{T}ext-to-\textbf{D}ecision \textbf{A}gent (\textbf{T2DA}), a simple and scalable framework that supervises offline meta-RL with natural language. We first introduce a generalized world model to encode multi-task decision data into a dynamics-aware embedding space. Then, inspired by CLIP, we predict which textual description goes with which decision embedding, effectively bridging their semantic gap via contrastive language-decision pre-training and aligning the text embeddings to comprehend the environment dynamics. After training the text-conditioned generalist policy, the agent can directly realize zero-shot text-to-decision generation in response to language instructions. Comprehensive experiments on MuJoCo and Meta-World benchmarks show that T2DA facilitates high-capacity zero-shot generalization and outperforms various types of baselines. Our code is available at \textcolor{magenta}{\href{https://github.com/NJU-RL/T2DA}{https://github.com/NJU-RL/T2DA}}.

cs.AI↗