Searcharxiv⌕ Search

arXiv subjects

Jiayi Cheng

Publications and source records attributed to Jiayi Cheng.

15 recordsLinked to original sources
arXiv2026

What Does Multi-Harness RL Learn? Credit Assignment and Portability in Coding Agents

Agent reinforcement learning (RL) increasingly runs through full execution harnesses, and a multi-harness recipe mixes two choices: exposing the policy to several harnesses, and comparing their rewards inside one relative-advantage group. We isolate the second choice in repository-level coding. From one Qwen3-8B supervised warm start we replay the same frozen task-harness records from Aider, OpenHands, Qwen Code, and SWE-agent, with the same number of updates, under two rules for group-relative policy optimization (GRPO), Within (one group per task-harness pair) and Cross (harnesses pooled within a task), and score every checkpoint with a sealed SWE-bench Verified oracle on four source harnesses and a minimal harness held out of training. The evaluation harness is the dominant variable: across 24,000 sealed evaluations it moves the mean solve rate from 2.14\% to 9.27\%, a factor of 4.3, where the training recipe moves it by 1.16. The grouping rule is not. On the held-out harness, Cross minus Within is +0.25 pp, 95\% confidence interval [-0.48, +1.02], at eight attempts per task, and +0.16 [-0.41, +0.72] pooled over three training seeds whose individual estimates change sign. Each rule's own seed range, 0.42 to 0.45 pp, exceeds the difference between them. Both rules place their largest gains on the same source harness. The pooled advantage carries the harness: an out-of-fold classifier recovers the generating harness from Cross's advantage +4.48 pp above the shuffled-label baseline and from Within's not at all, and the two rules still reach the same held-out score and action distribution inside each harness. Re-collecting half the training data on-policy does not change this. Cross-harness credit yields configuration adaptation and no more portable capability than within-harness credit. Multi-harness RL reports should state the grouping boundary and test under an unseen harness.

cs.AI↗
arXiv2026

PreDiff-LM: Pretrained Discrete Masked Diffusion Language Modeling with Hybrid Attention

Discrete masked diffusion language models support bidirectional generation and infilling, but adapting pretrained autoregressive (AR) transformers requires reconciling causal pretraining with bidirectional denoising. We study this problem at the level of attention rather than claiming AR-weight reuse itself as novel. PreDiff-LM preserves causal attention within the observed prompt while allowing full bidirectional attention within the masked target. Under a matched GPT-2 Medium, WikiText-103, 90K-step setup, this hybrid mask improves unconditional perplexity from 34.1 to 28.7 and MAUVE from 0.71 to 0.78 over uniform bidirectional attention with the same AR initialization. Attention adaptation also composes with a DiffuGPT-style objective adaptation, reaching 26.9 perplexity. Pretrained initialization reduces the steps required to reach perplexity below 50 from about 350K to 8K, although a compute-matched fine-tuned AR model remains stronger at equal scale (18.9 versus 28.7). Beyond perplexity, PreDiff-LM improves repetition, distributional quality, four zero-shot downstream tasks, and human preference over prior diffusion baselines. The results position hybrid attention as a complementary mechanism for adapting pretrained causal backbones, while making explicit the remaining quality and inference-efficiency gaps to optimized AR models.

cs.AI↗
arXiv2026

Less Data, Better Alignment: Data-Centric Multi-Evaluator Agreement for Preference Optimization

Research on preference optimization often varies the training objective while holding the data fixed. We instead ask whether a small, high-confidence set of on-policy responses can provide a reliable learning signal. Our method, DMAPO (Data-centric Multi-evaluator Agreement for Preference Optimization), generates candidate responses from the target policy, evaluates helpfulness, factuality, and conciseness with rubric-specialized evaluators, applies a process-critic correction, and retains only high-consensus desirable or undesirable examples. This procedure accepts 1,871 of 54,236 Mistral-7B candidates (3.45%). KTO trained on this set reaches 7.50 on MT-Bench, 95.5% length-controlled win rate against a text-davinci-003 reference, and 57.3% IFEval prompt accuracy. Independent pairwise evaluation also favors DMAPO over SimPO: GPT-4o yields a net win rate of 23.3 points on 129 held-out prompts and 24.0 points on 200 out-of-distribution LMSYS-Chat prompts; Claude Opus 4.7 yields 24.1 points on the held-out set. Changing the evaluator model or rubric alters the selected examples but has little effect on downstream performance. A second-backbone study yields a similar 3.41% acceptance rate, although its performance gains are more modest. Across these experiments, consensus filtering offers a data-efficient route to preference optimization for general instructions, at the cost of additional curation compute and dependence on evaluator judgments.

cs.AI↗
arXiv2026

CodeRescue: Budget-Calibrated Recovery Routing for Coding Agents

Coding agents increasingly operate in executable environments where a failed attempt produces actionable feedback rather than merely an incorrect answer. Existing cost-aware systems typically treat such failures as cascade decisions: try a cheap model first, then escalate hard cases to a stronger and more expensive model. In coding, however, execution feedback can also make further cheap-model recovery worthwhile, raising a budgeted deployment question: when should an agent spend more cheap compute, and when should it escalate? We formulate this post-failure decision as recovery routing over heterogeneous actions and train a supervised router from execution rollouts. To make the same router usable under changing budgets, we add a Conformal Risk Control (CRC) layer that selects a deployment-time cost penalty without retraining and provides marginal expected-cost control under exchangeability. Across held-out failures from five coding benchmarks, cheap recovery and escalation exhibit complementary success patterns. The calibrated frontier improves over fixed actions, prompt-only routers, and a binary cascade baseline; in the main GPT-5.4-nano/GPT-5.4 setting, one CRC-calibrated frontier point exceeds always-escalate solve rate while using 35% of its mean recovery cost. Code is available at https://github.com/Qijia-He/agent-budget-control.

cs.AI↗
arXiv2026

A Few Teacher Steps Go a Long Way: Cost-Efficient On-Policy Data Augmentation for Agent Post-Training

For LLM agents, supervised fine-tuning is not only about teacher labels' quality, but also about which interaction contexts those labels condition on. Pure behavioral cloning uses full teacher demonstrations, creating a mismatch between teacher-induced contexts seen in training and student-induced contexts encountered at test time. Recent work addresses this mismatch by querying a teacher at contexts reached by the student, often with increasingly elaborate filtering of the teacher's continuations. We instead frame on-policy data construction as a budget-allocation problem: under matched supervision resources, should teacher output be spent on more start-to-finish demos, longer continuations, outcome filtering, or broader coverage of learner-induced contexts? We formalize this design space through the rollout policy, switch-time distribution, continuation horizon, filtering rules, and two complementary costs: teacher inference generated before filtering and teacher supervision retained for SFT. Across HotpotQA, ALFWorld, and Terminal-Bench-Dev, bounded unfiltered teacher continuations at learner-induced contexts improve over pure behavioral cloning at matched budgets. On HotpotQA and ALFWorld, where we run the full comparison, few-step continuations match or exceed success-filtered and critical-context-filtered alternatives. Our findings suggest that a few teacher steps, placed at learner-induced contexts, can be a more cost-efficient supervision allocation than longer or more heavily curated teacher completions.

cs.LG↗
arXiv2026

Agents' Last Exam

Recent AI systems have achieved strong results on a wide range of benchmarks, yet these gains have not translated into economically meaningful deployment across many professional domains. We argue that this gap is largely an evaluation problem: widely used benchmarks lack sustained performance measurement on real and economically valuable workflows. This paper introduces Agents' Last Exam (ALE), a benchmark designed to evaluate AI agents on long horizon, economically valuable, real world tasks with verifiable outcomes. Developed in collaboration with 250+ industry experts, ALE covers non-physical industries defined with reference to O*NET / SOC 2018 (the U.S. federal occupational taxonomy). It is organized around a task taxonomy with 55 sub fields grouped into 13 industry clusters covering 1K+ tasks. Current results show that the hardest tier remains far from saturated: across mainstream harness and backbone configurations, the average full pass rate is below 1%. ALE is designed as a living benchmark: its task pool grows continuously as new workflows and industries are onboarded. More broadly, ALE is intended not merely as another leaderboard, but as an instrument for closing the gap between benchmark success and GDP relevant impact.

cs.AI↗
arXiv2026

Discrete Stochastic Localization for Non-autoregressive Generation

Continuous diffusion is a natural framework for non-autoregressive generation but has generally lagged behind masked discrete diffusion models (MDMs) on discrete sequence generation. We argue that the bottleneck is not continuity itself, but a representation in which denoising depends on timestep-indexed noise regimes. We introduce \emph{Discrete Stochastic Localization} (DSL), a continuous-state framework with unit-sphere token embeddings whose Bayes-optimal denoiser is invariant to the nominal signal-to-noise ratio (SNR) under the localization channel. One trained network then supports an entire family of per-token SNR paths, with endpoint masked-diffusion paths as a special case. Fine-tuning a pretrained MDLM checkpoint with DSL substantially improves distributional faithfulness (MAUVE) on OpenWebText across all step budgets from $T{=}128$ to $T{=}1024$, and the same checkpoint supports random-order autoregressive sampling, as well as a hybrid continuous-then-discrete sampler using as few as T=48 total steps -- without distillation or retraining.

cs.LG↗
arXiv2026

Discrete Stochastic Localization for Non-autoregressive Generation

Continuous diffusion is a natural framework for non-autoregressive generation but has generally lagged behind masked discrete diffusion models (MDMs) on discrete sequence generation. We argue that the bottleneck is not continuity itself, but a representation in which denoising depends on timestep-indexed noise regimes. We introduce \emph{Discrete Stochastic Localization} (DSL), a continuous-state framework with unit-sphere token embeddings whose Bayes-optimal denoiser is invariant to the nominal signal-to-noise ratio (SNR) under the localization channel. One trained network then supports an entire family of per-token SNR paths, with endpoint masked-diffusion paths as a special case. Fine-tuning a pretrained MDLM checkpoint with DSL substantially improves distributional faithfulness (MAUVE) on OpenWebText across all step budgets from $T{=}128$ to $T{=}1024$, and the same checkpoint supports random-order autoregressive sampling, as well as a hybrid continuous-then-discrete sampler using as few as T=48 total steps -- without distillation or retraining.

cs.LG↗
arXiv2026

VehAnchor: Metadata-Free Metric Scale Recovery from Vehicle Cues in Aerial Imagery

Autonomous aerial robots operating in GPS-denied or communication-degraded environments frequently lose access to camera metadata and telemetry, leaving onboard perception systems unable to recover the absolute metric scale of the scene. As LLM/VLM-based planners are increasingly adopted as high-level agents for embodied systems, their ability to reason about physical dimensions becomes safety-critical -- yet our experiments show that five state-of-the-art VLMs suffer from \emph{spatial scale hallucinations}, with median area estimation errors exceeding 50\%. We propose VehAnchor, a lightweight, deterministic \emph{Geometric Perception Skill} designed as a callable tool that any LLM-based agent can invoke to recover Ground Sample Distance (GSD) from ubiquitous environmental anchors: small vehicles detected via oriented bounding boxes, whose modal pixel length is robustly estimated through kernel density estimation and converted to GSD using a pre-calibrated reference length. The tool returns both a GSD estimate and a composite confidence score, enabling the calling agent to autonomously decide whether to trust the measurement or fall back to alternative strategies. On the DOTA~v1.5 benchmark, VehAnchor achieves 6.87\% median GSD error on 306~images. Integrated with SAM-based segmentation for downstream area measurement, the pipeline yields 19.7\% median error on a 100-entry benchmark -- with 2.6$\times$ lower category dependence and 4$\times$ fewer catastrophic failures than the best VLM baseline -- demonstrating that equipping agents with deterministic geometric tools is essential for safe autonomous spatial reasoning.

cs.RO↗
arXiv2026

FutureX-Pro: Extending Future Prediction to High-Value Vertical Domains

Building upon FutureX, which established a live benchmark for general-purpose future prediction, this report introduces FutureX-Pro, including FutureX-Finance, FutureX-Retail, FutureX-PublicHealth, FutureX-NaturalDisaster, and FutureX-Search. These together form a specialized framework extending agentic future prediction to high-value vertical domains. While generalist agents demonstrate proficiency in open-domain search, their reliability in capital-intensive and safety-critical sectors remains under-explored. FutureX-Pro targets four economically and socially pivotal verticals: Finance, Retail, Public Health, and Natural Disaster. We benchmark agentic Large Language Models (LLMs) on entry-level yet foundational prediction tasks -- ranging from forecasting market indicators and supply chain demands to tracking epidemic trends and natural disasters. By adapting the contamination-free, live-evaluation pipeline of FutureX, we assess whether current State-of-the-Art (SOTA) agentic LLMs possess the domain grounding necessary for industrial deployment. Our findings reveal the performance gap between generalist reasoning and the precision required for high-value vertical applications.

cs.AI↗
arXiv2025

Bayesian Distributionally Robust Merton Problem with Nonlinear Wasserstein Projections

We revisit Merton's continuous-time portfolio selection through a data-driven, distributionally robust lens. Our aim is to tap the benefits of frequent trading over short horizons while acknowledging that drift is hard to pin down, whereas volatility can be screened using realized or implied measures for appropriately selected assets. Rather than time-rectangular distributional robust control -- which replenishes adversarial power at every instant and induces over-pessimism -- we place a single ambiguity set on the drift prior within a Bayesian Merton model. This prior-level ambiguity preserves learning and tractability: a minimax swap reduces the robust control to optimizing a nonlinear functional of the prior, enabling Karatzas and Zhao \cite{KZ98}-type's closed-form evaluation for each candidate prior. We then characterize small-radius worst-case priors under Wasserstein uncertainty via an explicit asymptotically optimal pushforward of the nominal prior, and we calibrate the ambiguity radius through a nonlinear Wasserstein projection tailored to the Merton functional. Synthetic and real-data studies demonstrate reduced pessimism relative to DRC and improved performance over myopic DRO-Markowitz under frequent rebalancing.

math.OC↗
arXiv2025

MC3D-AD: A Unified Geometry-aware Reconstruction Model for Multi-category 3D Anomaly Detection

3D Anomaly Detection (AD) is a promising means of controlling the quality of manufactured products. However, existing methods typically require carefully training a task-specific model for each category independently, leading to high cost, low efficiency, and weak generalization. Therefore, this paper presents a novel unified model for Multi-Category 3D Anomaly Detection (MC3D-AD) that aims to utilize both local and global geometry-aware information to reconstruct normal representations of all categories. First, to learn robust and generalized features of different categories, we propose an adaptive geometry-aware masked attention module that extracts geometry variation information to guide mask attention. Then, we introduce a local geometry-aware encoder reinforced by the improved mask attention to encode group-level feature tokens. Finally, we design a global query decoder that utilizes point cloud position embeddings to improve the decoding process and reconstruction ability. This leads to local and global geometry-aware reconstructed feature tokens for the AD task. MC3D-AD is evaluated on two publicly available Real3D-AD and Anomaly-ShapeNet datasets, and exhibits significant superiority over current state-of-the-art single-category methods, achieving 3.1\% and 9.3\% improvement in object-level AUROC over Real3D-AD and Anomaly-ShapeNet, respectively. The code is available at https://github.com/iCAN-SZU/MC3D-AD.

cs.CV↗
arXiv2025

Duality and Policy Evaluation in Distributionally Robust Bayesian Diffusion Control

We study diffusion control problems under parameter uncertainty. Controllers based on plug-in estimation can be brittle due to potential distribution shifts. Bayesian control with a prior on the parameters offers a formulation with beliefs about such shifts. However, as with any Bayesian model, the prior may be misspecified. To mitigate misspecification and reduce over-pessimism compared to classical robust control approaches (e.g. \citet{hansen2008robustness}), we propose a distributionally robust Bayesian control (DRBC) formulation in which an adversary perturbs the prior within a divergence neighborhood of a baseline prior. We develop a strong duality result that reduces the distributionally robust prior evaluation to a low-dimensional optimization and yields a practical simulation-based policy evaluation and learning procedure with structured policy parameterizations. We validate the efficiency of the algorithm on a synthetic linear-quadratic control example and real-data portfolio selection.

math.OC↗
arXiv2024

Infrared spectroscopy study of kagome material CsTi$_3$Bi$_5$

The kagome material CsTi$_3$Bi$_5$, which is isostructural to the extensively studied charge density wave (CDW) compound CsV$_3$Sb$_5$, exhibits intriguing electronic features within its two-dimensional kagome lattices of titanium atoms. Here, we perform optical spectroscopic measurements together with the first-principles calculations on single-crystalline CsTi$_3$Bi$_5$ to investigate its electronic properties comprehensively. It is found that the overall optical spectra are very similar to those of CsV$_3$Sb$_5$, but the existence of CDW instability is ruled out in CsTi$_3$Bi$_5$. Via careful comparison to the optical responses of CsV$_3$Sb$_5$, we attribute this difference to a significant reduction in the itinerant carrier density of CsTi$_3$Bi$_5$, which is associated with the absence of van Hove singularity near the Fermi level at $M$ point. This result supports the scenario that the CDW in CsV$_3$Sb$_5$ is driven by the nesting of van Hove singularity. Additionally, we unveil some exotic low-lying absorption features, which provide clear evidence of flat bands in CsTi$_3$Bi$_5$. Our findings contribute to a deeper understanding of exotic phenomena in CsTi$_3$Bi$_5$ and provide valuable insights into the role of van Hove singularity in CsV$_3$Sb$_5$.

cond-mat.str-el↗
arXiv2023

SMPL: Simulated Industrial Manufacturing and Process Control Learning Environments

Traditional biological and pharmaceutical manufacturing plants are controlled by human workers or pre-defined thresholds. Modernized factories have advanced process control algorithms such as model predictive control (MPC). However, there is little exploration of applying deep reinforcement learning to control manufacturing plants. One of the reasons is the lack of high fidelity simulations and standard APIs for benchmarking. To bridge this gap, we develop an easy-to-use library that includes five high-fidelity simulation environments: BeerFMTEnv, ReactorEnv, AtropineEnv, PenSimEnv and mAbEnv, which cover a wide range of manufacturing processes. We build these environments on published dynamics models. Furthermore, we benchmark online and offline, model-based and model-free reinforcement learning algorithms for comparisons of follow-up research.

cs.LG↗