arXiv subjects

Yu Cheng

Publications and source records attributed to Yu Cheng.

At least 37 records · Page 2Linked to original sources

arXiv2026

MIRA: A Musical Intent Refinement Agent for Aligning Text-to-Music Generation with User Intent

Text-to-music systems produce increasingly convincing audio, yet evaluation reveals little about whether the result matches user intent. A global text-audio relevance score can overlook the implicit intent in underspecified prompts and mask failures in specific requirements, such as instrumentation, structure, rhythm, or mood progression. To bridge this gap, we formulate text-to-music intent alignment as satisfying a per-request rubric of independently verifiable items covering both a request's explicit requirements and its implied musical intent. Scoring items individually makes evaluation diagnostic by intent source and musical dimension, rather than a single opaque score. We instantiate this as MuRA-Bench, a benchmark of real-world platform requests curated by music experts. We further propose MIRA (Musical Intent Refinement Agent), a test-time agent that first grounds a request's intent into rubrics, then searches over prompt revisions for a black-box generator under a bounded budget, iteratively generating music, verifying it against the rubrics, and using this feedback to guide a trajectory-aware tree search. Experiments across open-source and commercial backends show that MIRA improves intent alignment, enabling an open-source generator to achieve performance comparable to representative commercial systems (e.g. Suno and Mureka). Project page: https://mirareview.github.io/.

cs.SD↗

arXiv2026

SubtleMemory: A Benchmark for Fine-Grained Relational Memory Discrimination in Long-Horizon AI Agents

Persistent AI assistants, such as OpenClaw, accumulate large collections of related memories over long-term interactions. As these memories grow, they may reinforce one another, diverge across contexts, or directly conflict, making correct assistance depend on memory relations rather than isolated recall. Existing long-term memory benchmarks do not systematically probe how agents preserve and utilize such relations during downstream tasks. To address this gap, we introduce SubtleMemory, a benchmark for fine-grained relational memory discrimination in long-running AI agents. SubtleMemory constructs relation-controlled latent semantic artifacts whose variants instantiate complementary, nuanced, or contradictory relations, and embeds them into realistic user-agent histories, requiring agents to recover distributed relational structures during later queries and instructions. The benchmark contains 1,522 evaluation instances over 10 long histories, grounded in 1,090 relation-controlled memory-variant sets and spanning user-related and non-user-related queries. Evaluating six standalone memory systems, two Claw-style agents with native memory modules, and three Claw-style agents with plugin memory modules, we find that current systems remain weak on fine-grained relational memory discrimination. We further introduce diagnostic protocols that reveal distinct capability profiles across memory preservation, retrieval, and downstream reasoning stages.

cs.AI↗

arXiv2026

Internalizing Agent Experience into Diffusion Model Weights via On-Policy Context Distillation

Wrapping an image generation model in an agentic harness can effectively boost Text-to-Image task performance: the harness can leverage memory, skills, workflow orchestration, result verification, and iterative refinement to continually construct and revise prompts, thereby eliciting better images. These gains, however, remain external to the diffusion model and are realized only while the full harness runs. We propose Diffusion On-Policy Context Distillation (D-OPCD), which treats the agent-improved prompt as privileged context and distills the knowledge encoded in the agent harness into the weights of the diffusion model, so that the model retains part of the harness's benefit when conditioned on the original query alone. Using a Text-to-Image agent equipped with our proposed Auto Skill Evolver (ASE), we show that D-OPCD can internalize harness capabilities into the generator's weights, raising the average direct-generation score from 60.52 to 65.09 across four benchmarks. With this knowledge absorbed into the weights, the harness can shed its saturated skills and resume evolving: a second ASE round on the updated generator improves on a skill-free harness by additional 1.83 points, pointing toward text-to-image systems in which harness and model keep improving each other through continual co-evolution.

cs.AI↗

arXiv2026

SafeCoEvo: Co-Evolving Safety Harnesses and Guards for LLM Agents at Test-Time

LLM agents deployed in real-world environments continually encounter new tasks and safety risks, while execution feedback typically becomes available only after each task is completed. However, existing self-evolving approaches commonly rely on multiple rounds of optimization over fixed and repeatedly accessible task distributions, fundamentally differing from test-time adaptation in real-world deployment, where only experience accumulated from past tasks can be used to improve safety decisions on future unseen tasks. To address this limitation, we propose SafeCoEvo, a test-time Harness-Guard co-evolution framework for LLM agent safety that enables the external safety system to continually adapt from accumulated runtime experience. SafeCoEvo jointly improves two complementary safety capabilities at different timescales: S-Harness rapidly externalizes recent runtime experience into updatable explicit safety knowledge that can promptly influence subsequent tasks, while GuardVPO internalizes accumulated runtime safety experience over a longer timescale into parametric risk-judgment capabilities. By combining short-term rapid adaptation with long-term capability consolidation, SafeCoEvo continually improves the agent's safety capabilities, reducing the unsafe outcome rate by 10.05% while improving the task success rate by 12.15% over the strongest baseline, thereby achieving simultaneous gains in safety and task utility.

cs.AI↗

arXiv2026

SERA: Scale-Equalized Rollout Allocation for Maximum Likelihood Reinforcement Learning

Maximum Likelihood Reinforcement Learning (MaxRL) targets prompt-wise log-success and has shown strong performance on reasoning tasks. Under finite rollout budgets, however, the estimator used by MaxRL attenuates each prompt's likelihood gradient by a factor that depends on its success probability and rollout count. Under uniform rollout allocation, the common rollout count fails to compensate for success-dependent attenuation, leaving low-success prompts more strongly attenuated and distorting their relative contributions to the expected aggregate gradient. We introduce SERA (Scale-Equalized Rollout Allocation), which redistributes a fixed rollout budget to approximately equalize these finite-rollout scaling factors. Building on our theoretical analysis of how finite rollouts distort prompt-wise likelihood gradients, we formulate the allocation as a fixed-budget max--min problem, derive a waterline solution to its continuous relaxation, and introduce a multiplicity correction to remove the additional prompt weighting induced by heterogeneous rollout counts. Experiments show stronger alignment with exact likelihood gradients in a controlled ImageNet setting and improved multi-sample solution coverage over MaxRL on maze navigation and mathematical reasoning under matched training rollout budgets.

cs.LG↗

arXiv2026

TMCS: Tool-Grounded Multi-Agent Reasoning for Compositional Chemical Problem Solving

Despite the promise of Large Language Models (LLMs) in computational chemistry, rigorous combinatorial chemistry problems remain difficult because they require quantitatively constrained molecular modification, candidate validation, and systematic revision after failed attempts. Existing tool-augmented chemical agents demonstrate useful planning and tool use, but they rarely provide a unified loop for property-driven molecular optimization and workflow-level composition. To bridge this gap, we propose Tool-Grounded Multi-Agent Reasoning for Compositional Chemical Problem Solving (TMCS), a step-by-step multi-agent framework that formalizes chemical problem solving as an interpretable, tool-augmented workflow. At the task level, specialized agents leverage external tools, few-shot trajectory memory, and structured reflection to iteratively refine solutions. At the workflow level, TMCS chains generation, understanding, editing, description, and optimization into a closed-loop pipeline. Evaluations across multiple chemical tasks demonstrate that TMCS consistently enhances chemical reasoning across both open- and closed-source base models, achieving state-of-the-art performance.

cs.AI↗

arXiv2026

HiFloat4 Format for Language Model Pre-training on Ascend NPUs

Training large foundation models at low numerical precision is one of the most promising directions for reducing the compute and memory cost of modern AI. Recent 4-bit floating-point formats such as MXFP4 and NVFP4 can be applied to linear GEMM operations in LLMs, but their limited dynamic range introduces numerical instability that prior work addresses by stacking stabilization mechanisms, typically executed at higher precision and partially eroding the efficiency gains that motivate FP4. In this work, we argue that numerical format design is itself a first-class lever for stable FP4 training, and present the first systematic study of FP4 LLM pretraining on energy-efficient Huawei Ascend NPUs. We compare the recently proposed HiFloat4 (HiF4) format against both MXFP4 and NVFP4, holding one recipe fixed across all three formats across dense (OpenPangu-1B, Llama3-8B) and Mixture-of-Experts (Qwen3-MoE-30B) architectures and executing all linear and expert GEMMs in FP4. At matched storage --- NVFP4 and HiF4 both spend 4.5 bits per value --- the three formats differ far more in what they require before they will train at all than in final accuracy. HiF4 reaches a relative loss of 1.55\% with no stabilization, below fully stabilized MXFP4 (1.79\%) and below NVFP4 carrying the per-tensor scaling it cannot train without (2.00\%); NVFP4 diverges under every combination of stochastic rounding and Hadamard transform we tried. Our results suggest that stable, accurate FP4 training does not require an ever-growing stack of stabilization techniques; it requires the right numerical format.

cs.LG↗

arXiv2026

SciWalker: Synthesizing Scientific Coding Problems with Operator Graphs and Execution Feedback

Improving the scientific coding capabilities of large language models (LLMs) requires high-quality training data. However, such data remain scarce because manually authoring realistic problems is costly and time-consuming, while systematically covering diverse scientific domains and algorithmic combinations remains challenging. To address this, we introduce SciWalker, a framework for synthesizing scientific coding problems through operator-chain sampling and execution feedback. The framework combines scientific library interfaces with operation modes to instantiate operators, organizes them into operator graphs, and samples operator chains as computational workflow cues. Guided by these cues, we adopt LLMs to generate scientifically grounded problem statements, reference solutions, and tests, with failed generations iteratively repaired using execution feedback. By combining structured workflow composition with verification and quality review, SciWalker enables scalable task generation while promoting scientific grounding, computational diversity, and executability. Using this framework, we construct 8,178 high-quality problems spanning 5 scientific domains and 32 subdomains. To evaluate their training utility, we conduct reinforcement learning on Qwen3.5-9B using the GSPO algorithm. This training improves SciCode subproblem accuracy by 9.9 percentage points, from 29.3% to 39.2%, with gains across scientific code generation, code repair, and reasoning benchmarks. The code for SciWalker is available at https://github.com/lichenx1/SciWalker.

cs.AI↗

arXiv2026

The Past Frames the Future: Memory for Autoregressive Video Generation

Advances in generative models have improved video fidelity, enabling long-horizon generation, interactive world modeling, and evolving visual environments. Autoregressive (AR) video generation extends visual sequences through causal rollouts. However, a fundamental bottleneck emerges: as the generated sequence expands, practical models must operate under strictly bounded context windows, storage, and computational limits. Consequently, critical historical information, e.g., entity identities, dynamic states, and intervention-induced causal changes, often leaves the active context long before its relevance diminishes. Overcoming this limitation and maintaining temporal persistence constitutes a fundamental memory problem. We present a systematic and comprehensive review of memory mechanisms in AR video generation. We formulate memory operationally as persistent historical information maintained across outer AR steps, capable of influencing future generation even after the originating evidence is no longer locally accessible. Building upon this unified framework, we organize the literature through five complementary perspectives: (I) Forms, the representational carriers of history; (II) Functions, the specific semantic and physical information requiring preservation; (III) Operations, the lifecycle of writing, reading, updating, managing, and integrating memory; (IV) Learning, the optimization of memory behaviors under closed-loop rollouts; and (V) Evaluation, the paradigms for diagnosing genuine memory capabilities. We conclude by synthesizing open challenges, including composable and resource-aware memory architectures, trustworthy state updating, self-rollout learning, and standardized evaluation. By bridging representations, mechanisms, and learning paradigms, this paper establishes a structured foundation for developing reliable, memory-conditioned video generation systems.

cs.CV↗

arXiv2026

Low-Metallicity Star Formation Survey in Sh2-284 (LZ-STAR): The Core Mass Function

We present an ALMA 1.3~mm dust continuum study of the dense core mass function (CMF) in Sh2-284, a low metallicity outer Galaxy star-forming complex with $Z\sim1/3$--$1/2~Z_\odot$. The observations cover six far-infrared bright subregions at $\sim$0.65" (3000~au) resolution. We identify a total of 91 candidate dense cores and define a robust catalog of 68 cores with Gaussian fitting. The high-mass CMF above 2.5 Msun is well described by a Salpeter-like slope, with a fiducial forward-modeled value of $α=1.22^{+0.24}_{-0.22}$ for $dN/d\log M\propto M^{-α}$. Together with existing constraints on the initial mass function (IMF) of Sh2-284, the Salpeter-like CMF is consistent with a resemblance between the CMF and IMF shapes in the outer Galaxy environments, suggesting that moderately low metallicity alone does not strongly reshape the high-mass CMF/IMF slope.

astro-ph.GA↗

arXiv2026

HySparse2: Hybrid Sparse Attention with Two-Level KV Sharing

Long-horizon and multi-turn agents typically generate short actions and process long observations from tools and environments. This growing context demands efficient prefill, compact KV-cache storage, and accurate long-context retrieval. To meet these demands, we introduce HySparse2, a hybrid sparse attention architecture with two-level KV sharing. At the outer level, KV Bridging adopts a YOCO-style self-decoder and cross-decoder structure, but bridges only full-attention layers. The self-decoder uses hybrid sliding-window attention (SWA), while the cross-decoder uses hybrid sparse attention. The KV caches for full-attention layers in the cross-decoder are generated from the hidden states of full-attention layers in the self-decoder. At the inner level, HySparse2 retains HySparse's core KV Reuse design with two refinements. First, it replaces block-level sparsity with token-level sparsity for finer long-context retrieval. Second, it removes the separate SWA branch from sparse layers and instead forces a sliding window of recent tokens into the sparse selection. This two-level KV sharing allows all cross-decoder KV caches to be constructed from self-decoder hidden states. Prefill can therefore exit after the self-decoder, skipping all cross-decoder layers. On an 80B-A3B MoE model, HySparse2 outperforms HySparse and Hybrid SWA on long-context retrieval and multi-turn agentic tasks, while substantially reducing prefill computation and KV-cache storage.

cs.CL↗

arXiv2026

A High Resolution View of Three Massive Young Stellar Objects in the Magellanic Clouds

The Magellanic Clouds provide an excellent laboratory to study massive star formation beyond typical Galactic environments. Using ALMA band 7 long-baseline observations, we present the highest resolution (approximately 0.03", $1500{\rm\:au}$) view to date of three massive young stellar objects (MYSOs) in the Magellanic Clouds, i.e., Lh03 and Lh09 in the LMC, and S07 in the SMC. The 0.87 mm continuum, with a sensitivity of $\sim30\:μ$Jy beam$^{-1}$, reveals varying levels of fragmentation, from a single compact source (S07), a few (Lh03), to a cluster of compact sources (Lh09). In Lh09, we identified 17 core candidates with lowest measured mass of 2.6 Msun, and their de-projected separations peak around 6000~au. All three targets exhibit high velocity CO(3--2) outflows, including detection of a collimated bipolar jet driven by a core in Lh09. Spectral lines from HCO$^+$, SO, and SO$_2$ are detected in all sources. We observe velocity gradients that likely trace rotating toroids (Lh03) or more complex motions (Lh09, S07). Taken together, these phenomena observed on $\sim1500$~au scales are qualitatively similar to those seen in Galactic MYSOs, suggesting a degree of universality in the basic processes of massive star formation across different environments. Nevertheless, the substantial source-to-source diversity in continuum morphology and dense-gas kinematics motivates larger high-resolution surveys to determine how these properties vary with source characteristics and environment.

astro-ph.GA↗

arXiv2026

Draft-OPD: On-Policy Distillation for Speculative Draft Models

Speculative decoding accelerates large language model inference by pairing a target model with a lightweight draft model whose proposed tokens are verified in parallel. A common way to build draft models, like EAGLE3 or DFlash is supervised fine-tuning (SFT) on target-generated trajectories. However, we observe that SFT quickly plateaus: the draft model's acceptance length on test data stops improving. The reason is an offline-to-inference mismatch: In SFT, the drafter learns from fixed target-generated trajectories, whereas during speculative decoding it is evaluated on blocks proposed under its own policy. This motivates on-policy distillation (OPD), where the target model supervises the drafter on draft-induced states. Yet OPD remains difficult for draft models, as they cannot reliably roll out complete sequences independently, whereas target-assisted generation makes the collected sequences follow the target distribution and thus eliminates the on-policy signal. We therefore propose Draft-OPD, which uses target-assisted rollout for stable continuations and replays drafting from the verification-exposed error positions. This allows the drafter to learn from target feedback on both accepted and rejected proposals, focusing training on the draft-induced errors that limit speculative acceptance. Experiments show that Draft-OPD achieves over $5\times$ lossless acceleration for thinking models across diverse tasks, improving over EAGLE-3 and DFlash by 23\% and 13\%.

cs.CL↗

arXiv2026

Global Minimality of Matrix-Valued Heteroclinic Connections between Vacuum Orbits

We study global energy minimization for matrix-valued heteroclinic connections whose vacuum sets are continuous conjugacy orbits rather than isolated points. For two distinguished parameter regimes of a one-dimensional $\mathrm{SU}(5)\times\mathbb Z_2$ matrix field model, we first identify the full zero set of the quartic potential as the disjoint union of two compact homogeneous vacuum orbits. We then construct auxiliary potentials controlled by the Frobenius distance to these orbits. Pointwise comparison with the original potential, combined with a calibration inequality and the one-dimensional coarea formula, gives a sharp lower bound for every finite-energy connection joining the two vacuum components; no diagonality, commutativity, or symmetry-reduced ansatz is imposed on the competitors. Two explicit non-Abelian kinks attain the bound and are therefore global minimizers in the full orbit-to-orbit heteroclinic classes. Their exact wall tensions are $4\sqrt2\,μ^3/(3λ)$ and $9\sqrt2\,μ^3/(10h)$, respectively. Equivalently, the corresponding line-segment paths realize the degenerate weighted distance between the vacuum manifolds. We also resolve every fixed-endpoint sector: for arbitrary prescribed representatives on the two vacuum orbits, the energy infimum is the same wall tension, and it is attained if and only if the prescribed representatives form a closest pair of the two orbits. Nonclosest endpoint sectors therefore exhibit non-attainment through increasingly slow tangential motion along the vacuum manifolds. The proof separates a model-independent distance--coarea criterion from the model-specific orbit geometry and polynomial positivity estimates, providing a reusable mechanism for matrix-valued multiwell potentials with symmetry-generated vacua.

math.AP↗

arXiv2026

Dust Seeding Molecules in a Massive Protostar -- Detection of TiO in Orion Source I

We report the first detection of TiO in star-forming regions based on Atacama Large Millimeter/submillimeter Array observations of Orion Source I, a well-characterized massive protostar. Multiple rotational transitions are identified, with emission spatially resolved within $\sim 50$ au, showing a compact distribution with a velocity structure consistent with the base of a rotating outflow. The spatial and velocity distributions of TiO are consistent with those of AlO, with both species being key dust seeding refractory molecules. The column density of TiO is derived to be $(3.0 \pm 0.4)\times10^{15}\ {\rm cm^{-2}}$, corresponding to $X_{\rm TiO/SiO} \sim 12.8 \pm 1.7 \times10^{-3}$, higher than CI chondrites and indicative of efficient dust-to-gas conversion near the protostar. The detection of TiO, a key dust seeding species, offers important constraints on refractory chemistry and the formation environments of primitive minerals, linking astrochemical processes in protostellar systems to the earliest stages of Solar System material formation.

astro-ph.GA↗

arXiv2026

Probing Feebly Interacting Particles with 511 keV Line from Circumstellar Medium of Supernovae

Core-collapse supernovae (CCSNe) are intense sources of feebly interacting particles (FIPs), whose visible decays beyond the stellar envelope can generate distinct electromagnetic signals. We propose the circumstellar medium (CSM) surrounding the CCSNe progenitor as a new target for 511 keV line searches. FIPs escaping the stellar interior may decay into electron-positron pairs and inject positrons into the CSM, where they slow down and subsequently annihilate, giving rise to a 511 keV line as an electromagnetic precursor to the supernova shock breakout. Using dark photons as a benchmark, we calculate the time-dependent signal for CSM profiles motivated by SN 2023ixf and SN 2024ggi. For a Galactic supernova at 10 kpc, the projected reach of COSI and AMEGO extends beyond existing supernova limits over part of the MeV-scale dark-photon parameter space. The same FIP decays can simultaneously heat the CSM and sublimate circumstellar dust, producing a rapid spectral transition from infrared excess to optical or ultraviolet emission [Y. Cheng et al., Phys. Rev. Lett. 137, 111005 (2026)]. The coincident appearance of this transition and the 511 keV line would provide a distinctive correlated multiwavelength signature of FIP decays in the dense presupernova environment.

hep-ph↗

arXiv2026

Rethinking Critic Learning in PPO: Understanding and Mitigating Value Flattening

In reinforcement learning for large language models, Proximal Policy Optimization (PPO) commonly uses a critic to estimate state values and reduce the variance of policy updates. However, we uncover a systematic failure mode in PPO critics, which we call Value Flattening: state values, estimated from multiple Monte Carlo continuations, change sharply across intermediate states while critic predictions remain comparatively flat. We further observe this phenomenon in a controlled FrozenLake environment and find that it becomes more pronounced as the state space grows. Our theoretical and empirical analyses relate Value Flattening to an implicit variance penalty in the critic loss and redundant updates from temporally correlated states with similar gradients. Motivated by these findings, we introduce SParse Proximal Policy Optimization (SP$^3$O), which applies the value loss to only a few well-separated states in each response to mitigate both effects. Experiments on Qwen3-Base show that SP$^3$O with only three states supervised per response can mitigate Value Flattening and consistently improve the learned policy across model sizes and evaluation suites. Together, our results identify Value Flattening as an important yet overlooked failure mode of critic learning in standard PPO and show that a simple sparse supervision strategy can mitigate it.

cs.LG↗

arXiv2026

Circumstellar Medium of Supernovae as New Probes for Feebly Interacting Particles

We propose a novel strategy to probe feebly interacting particles (FIPs) by exploiting the dense, confined circumstellar medium (CSM) surrounding core-collapse supernovae. FIPs produced in the protoneutron star can deposit substantial visible energy into the CSM via decay prior to the shock breakout from the progenitor star. This energy injection heats and ionizes the CSM, establishing an FIP-induced photosphere that generates distinctive precursor blackbody emission. Using early-time observations of supernova 2023ixf, we translate the nondetection of excessive precursor luminosity into stringent new constraints on MeV-scale dark photons as an exemplary model. Our results significantly extend existing core-collapse supernova bounds and exclude previously unexplored regions of parameter space. We further demonstrate that the FIP-induced dust sublimation offers key diagnostics for future Galactic supernovae, opening a new avenue to explore the dark sector.

hep-ph↗