arXiv subjects

Zhiyuan Yao

Publications and source records attributed to Zhiyuan Yao.

At least 19 recordsLinked to original sources

arXiv2026

The MAGE Project: Size Evolution and Disk Formation in Massive Galaxies at Cosmic Dawn

We investigate size evolution and disk formation with a new suite of eleven cosmological zoom-in simulations, Massive Assembly of Galaxies in the Early universe (MAGE), run with RAMSES at $\sim10\,\mathrm{pc}$ resolution until $z\simeq9$. The target galaxies span stellar masses of $M_\star\simeq2\times10^8$-$3\times10^{10}\,M_\odot$ at $z=9$. The stellar half-mass radius grows steeply with stellar mass at low masses, turns over near $M_\star\simeq5\times10^9\,M_\odot$, and evolves weakly with redshift. The intrinsic UV sizes in MAGE overlap the compact end of observed distributions, including the compact strong-line systems, but not the more extended population at $z>9$. Dust, PSF-limited size measurements, and our overdense environments may contribute to this offset. In the simulations, stars dominate the mass within the half-mass radius above $M_\star\simeq5\times10^8\,M_\odot$, where the central circular speed exceeds a characteristic stellar-feedback scale. Outflows are then increasingly retained within the galaxy, consistent with gas recycling on galactic rather than halo scales, which sustains central gas supply and drives a compaction-like central buildup. The turnover in size at $M_\star\simeq5\times10^9\,M_\odot$ coincides with rapid growth of rotational support to $v/σ\simeq1.5$-$3$ in star-forming gas and stars younger than 10Myr, as the central circular speed exceeds the halo virial speed, which the star-forming gas velocity dispersion tracks. The subsequent low central gas fraction, declining central star formation, and younger outskirts suggest the onset of inside-out quenching.

astro-ph.GA↗

arXiv2026

How Much Do Teacher Bridges Matter in Interleaved On-Policy Distillation?

On-policy distillation (OPD) trains a student using teacher predictions on student-generated responses. Some methods insert teacher-written segments, or bridges, but disagree on how to train on those tokens. We compare no loss, forward KL and reverse KL on identical trajectories using Qwen3-1.7B as student and Qwen3-4B-Instruct-2507 as teacher. We separate gradients by which model wrote each token. Under our short training runs and BF16 inference, pre-specified tests detected no added accuracy benefit from teacher text as context or as a training target. A one-paragraph bridge after the student's first paragraph is only 2.7% of the tokens and contributes a small gradient nearly orthogonal to the student-token gradient. In this main comparison, the three loss treatments gave similar gradient directions and were statistically equivalent in accuracy within a +/-5 percentage-point margin. An exploratory pilot placing the bridge where failed responses disagreed most with the teacher suggested a benefit from teacher-token loss. A pre-specified replication did not confirm this, and fourfold weighting showed no significant accuracy benefit. Two limitations bound these findings. Among the tested amounts of training on fixed responses, gains peaked after four updates before the student increasingly failed to stop generating. None of the tested online training variants met the criteria for stable stopping and improved accuracy. Our evaluations used BF16 inference, which rounds away most of the small weight changes. When evaluated in exact FP32, the tested replication models frequently reached the response-length limit. Our accuracy findings therefore concern BF16 inference; they do not establish that teacher-token supervision is generally unnecessary.

cs.LG↗

arXiv2026

Work Agents: Systems, Capabilities, and Optimization

As foundation models evolve from information generation to digital work execution, agents are increasingly capable of completing complex tasks across tools, applications, and workflows. However, existing research lacks a unified perspective on how agents reliably accomplish delegated work. To address this gap, we present a work-centric, system-level perspective that views work agents as systems combining foundation models and runtime harnesses to execute tasks in stateful workspaces and deliver verifiable outcomes. We introduce two complementary dimensions, work persistence and work openness, to characterize working conditions, and classify digital work into five forms according to task scope and organization. Building on this framework, we identify five core capability families required for reliable work execution and examine approaches to model optimization, harness optimization, and their joint optimization. We further analyze existing evaluation methods and metrics from the perspective of work outcomes. Finally, we discuss future research directions, including sustained work delegation, adaptation to open environments, and the interaction between work execution and capability improvement. This paper provides a unified framework for understanding, optimizing, and evaluating work agents, supporting the development of systems capable of delivering reliable, verifiable, and practically valuable work outcomes.

cs.LG↗

arXiv2026

Muse Glimmer 30B: An On-Device Agentic Model

We introduce Muse Glimmer, an open-source 30B parameter causal language model with perception encoder, purpose-built for autonomous agentic tasks on consumer hardware. Muse Glimmer integrates multi-step reasoning, reliable tool use and failure recovery into a single model that runs locally on consumer devices without requiring cloud infrastructure. The model accepts text and image inputs and produces text outputs, was trained on data from over a hundred languages, and is quantized to approximately 4-bit precision for efficient single-GPU inference. Muse Glimmer achieves state-of-the-art performance for its size class on agentic benchmarks, setting a new standard compared to the models of its size: Gemma4-31B and Qwen3.6-27B on tool use, multi-step planning, and agentic task completion. We release the model under the Apache 2.0 license to foster open-source development of capable, safe, local AI agents.

cs.CL↗

arXiv2026

Exact Chained-Word Threshold and Monotone-Path Structure in the Even Root-of-Unity Kaleidoscope Yang-Baxter Algebra

Qiu, Guan, and Yu conjectured that in the even-order Kaleidoscope Yang--Baxter equation every word $ΛT^{k_1}Λ\cdots T^{k_M}Λ$, built from a square-zero matrix $Λ$ and integer powers of a shift $T$ of order $N$, vanishes once $M\geq N/2$. We prove this for every even $N\geq4$ and every complex parameter $u$ with $u\neq0$ and $u^N\neq1$, and show that the bound is sharp. For $M=N/2-1$ we determine exactly which words are nonzero; each of them has rank two. After block reduction, the factors associated with the intervening powers of $T$ lower a common flag of length $N/2$ by at least one step. The spectral idempotents of $T$ then give a presentation of the algebra generated by $T$ and $Λ$ by paths between levels, in which two paths compose only when the level changes monotonically. This algebra is independent of $u$, has dimension $N^2-N+2$, and is a pullback of two incidence algebras. Its radical is generated by $Λ$ and has nilpotency index $N/2+1$, and the algebra has infinite global dimension and infinite representation type. Replacing powers of $T$ by arbitrary functions of $T$ gives exact rank formulas and, for $N\geq6$, rank-one elements in the top radical power. Finally, four conditions on a filtered pair of operators over an arbitrary field recover the threshold and the radical; none of them can be omitted.

math-ph↗

arXiv2026

Scaling Multi-Agent Intelligence with Substrate-Centric Coordination

Existing large language model (LLM) based multi-agent systems have been shown to scale poorly, and we identify three problems: reading burden that grows with the group, useful information diluted by shared work of unknown quality, and fragility to failed agents. We argue that the problems are largely due to agent-centric designs: the coordination structure is wired into fixed workflows, central coordinators, or peer-to-peer messaging, and recent blackboard-style methods that move the coordination into a shared medium relieve only the last problem. To close the gaps, we build ScaleMAS, where asynchronous agents interact with a shared datastore of linked messages via three operations: publishing messages with optional references to earlier ones that they build on, searching for messages by content, and visiting messages by address. Search returns at most $k$ messages, so an agent's read stays bounded at any group size; results are ranked by relevance and by how often they are cited, which we find to be an effective proxy for quality; and messages persist after their producers leave, so the group carries on when individual agents fail. Across four agentic tasks, ScaleMAS outperforms agent-centric and blackboard-style baselines at lower inference compute, continues to scale up to 128 agents, and tolerates losing agents midway.

cs.CL↗

arXiv2026

Travel Back To Move Forward: Retrospective Exploration and Evolving for Long Horizon Agents

Large language model agents increasingly operate in long-horizon environments where early mistakes can trigger cascading failures and trap agents in dead-end trajectories. Agents must therefore be able to proactively recover from such states and turn their exploratory experience into guidance for subsequent decisions. We introduce ReTRAIL, a framework that enables agents to rewind the environment to earlier states while preserving valid progress and carrying forward knowledge learned from abandoned exploration. The environment state moves backward while the agent's knowledge continues to move forward. ReTRAIL trains a shared policy with two complementary signals: State-Anchored GRPO compares alternative continuations from revisited states to improve execution, while a branch-advancement reward guides reflection through the progress of subsequent exploration. Experiments across four interactive benchmark MineSweeper, Sokoban, ALFWorld and ScienceWorld show that simply adding a rewind primitive provides limited benefit and can even degrade performance. In contrast, ReTRAIL enables agents to rewind from unproductive states and use insights from abandoned branches to guide subsequent exploration. These results demonstrate the value of coupling environment rewinding with experience reuse within an ongoing task for long-horizon task solving. The code is available at https://github.com/whi497/ReTRAIL.

cs.AI↗

arXiv2026

Cold Stream Penetration of Virial Shocks: Fragmentation, Coagulation, and Disruption in the Hot Circumgalactic Medium

Cold streams penetrating virial shocks of massive halos along cosmic web filaments are expected to fuel galaxy growth at high redshift, yet the physical processes governing their penetration remain uncertain. We investigate cylindrical cold streams penetrating a hot circumgalactic medium (CGM) using idealized three-dimensional simulations. We systematically vary the stream radius, Mach number, and initial stream-to-CGM pressure contrast across three density contrasts, while controlling stream properties after pressure equilibrium is re-established. We identify three evolutionary regimes: coagulation, fragmentation, and disruption, plus a borderline regime in which the stream core marginally survives while detached fragments are disrupted. At modest pressure contrast, survival depends primarily on the competition between velocity shear and radiative cooling. Increasing pressure contrast produces a transient response during pressure restoration, temporarily enhancing or suppressing cold-gas mass and cold-hot interfacial area before evolution converges to a shear-dominated state. At larger pressure contrasts, the oblique shock steepens into a bow shock, and the final outcome is determined by the ratio of post-shock cooling time to virial crossing time. In all survival cases, post-equilibration evolution is well described by turbulent radiative entrainment at the stream-CGM interface: cold-gas mass flux increases while mean streamwise momentum flux remains approximately conserved. Applying these results to a galaxy evolution framework, we find that cold streams are expected to penetrate virial shocks around massive halos at $z>2$, though they may fragment into multiphase structures within the hot CGM at $M_\mathrm{v}\gtrsim2\times10^{12}\,\mathrm{M}_\odot$. Toward lower redshift, bow-shock heating and inefficient post-shock cooling can suppress cold-stream penetration.

astro-ph.GA↗

arXiv2026

EnvACE: Internalizing Environment Dynamics via World Rehearsal for Agentic Reinforcement Learning

Training large language model agents for long-horizon tool use typically relies on interactions with real or synthesized executable environments, whose construction and verification are costly, or on external simulators that are difficult to ground. We introduce EnvACE, an agentic reinforcement learning method that replaces external environment interaction during training with world rehearsal. The policy alternates between acting and rehearsal: it first generates a tool call, then plays the role of the environment to produce the response induced by that action, and conditions subsequent decisions on the rehearsed response. Both roles are jointly optimized end-to-end using task-success rewards. Through world rehearsal, the policy internalizes the relationship between actions and their environment responses in its parameters, yielding an agent world model that directly supports decision making. Across BFCL-v4, tau^2-Bench, VitaBench, and FinMCP-Bench, EnvACE achieves strong and transferable performance, outperforming environment-scaling baselines in the overall evaluation. Controlled studies further show that world rehearsal consistently improves policy learning across model scales. At test time, the internalized world model enables private rehearsal before committed execution, yielding further gains under a moderate rehearsal budget without additional external interaction. Our findings establish world rehearsal as a new path toward scaling LLM agent training beyond the constraints of external environments. Our code is publicly available at https://github.com/Within-yao/EnvACE.

cs.AI↗

arXiv2026

Large Knowledge Model: A Knowledge Foundation for Agentic Science at Scale

Agentic science envisions many autonomous agents investigating concurrently while building on a shared, evolving body of scientific knowledge. This requires a knowledge foundation that supports high-concurrency access, preserves traceable and reusable reasoning, and grows incrementally. We propose the Large Knowledge Model (LKM), a growing, agent-native knowledge foundation that provides a general representation of scientific knowledge across disciplines. LKM organizes the scientific literature into reasoning graphs, with claims as the core nodes and associated reasoning chains that make explicit how premises and evidence support conclusions. These source-grounded objects are persistent and addressable; cross-paper links organize them into aligned question, workflow, and evidence views. Newly extracted papers extend the foundation incrementally while preserving existing object identities. Building on this foundation, we develop an agent-native, reasoning-aware scientific retrieval system that retrieves claims together with their reasoning chains and sources, enabling agents to inspect and reuse the evidence underlying scientific conclusions. Across benchmarks, agents using LKM retrieve more evidence, cite more faithfully, and answer scientific questions more accurately: LKM nearly doubles the known supporting and contradicting evidence retrieved on SciFact-Open (818 versus 443 claim-paper pairs), reasoning graphs raise citation F1 on ScholarQABench by more than 5 points over the same retrieved papers, and LKM retrieval improves a fixed answering model by 9.3, 4.2, and 14.7 points over no retrieval on ChemBench, PubMedQA, and SciBench. LKM lays the foundation for a scientific ecosystem in which AI scientists not only recall accumulated knowledge but also extend it, returning new questions, workflows, and evidence to a memory that every subsequent investigation can build on.

cs.AI↗

arXiv2026

When Agents See Differently: Exposing UI Desynchronization Threats in Mobile Agents

Mobile agents are increasingly capable of autonomously interacting with mobile applications and performing consequential actions on behalf of users. Effective human oversight of such agents relies on a basic premise: users and agents observe consistent information from the same interface. We show that this premise can be systematically violated. Users perceive mobile interfaces through physical displays and the human visual system, making their observations subject to occlusion and luminance contrast limitations. In contrast, agents consume digital screenshots that may retain such content and accessibility representations that expose nonvisual widget metadata. The same UI state can therefore present materially different information to users and agents, a mismatch we term human-agent UI desynchronization. We investigate whether a repackaged clone of a legitimate APK can exploit this desynchronization to steer an agent toward attacker-designated actions, while remaining fully functional and behaviorally consistent with the original application for human users. We demonstrate that this threat is feasible: perturbations embedded before deployment can induce such deviations without access to runtime user instructions, agent detection or online adaptation. To systematically expose and evaluate this threat, we develop an automated framework that constructs user runtime instruction-agnostic UI desynchronization attacks and realizes them in deployable APKs. We conduct static and dynamic evaluations across five mobile-agent frameworks and three backbone models on 546 tasks involving various applications, achieving average misleading rates of 77.9% and 66.9%, respectively. A complementary questionnaire-based study with 186 participants finds that the visual perturbations used in our attacks are difficult for human users to notice.

cs.CR↗

arXiv2026

From Feedback-Free Star Clusters to Little Red Dots via Compaction

We address the origin of the Little Red Dots (LRDs) seen by JWST at cosmic morning ($z \!=\! 4 \!-\! 8$) as compact stellar systems with over-massive black holes (BHs). We propose that LRDs form naturally after feedback-free starbursts (FFB) in thousands of star clusters and following wet compaction. Analytically, we show how the clusters enable efficient dry migration of stars and BHs to the galaxy center by two-body segregation and dynamical friction against the disk. The clusters merge to form compact central stellar systems as observed. Mutual tidal stripping does not qualitatively affect the analysis. The young, rotating clusters are natural sites for the formation of BH seeds via rapid core collapse. The migrating clusters carry the BH seeds, which merge into central super-massive BHs (SMBHs). Compactions are required to deepen the potential wells such that the SMBHs are retained after post-merger gravitational-wave recoils, locked to the galaxy centers. Using cosmological simulations at different epochs, with different codes and physical recipes, we evaluate the additional growth of LRD-matching compact central stellar systems by global compaction events. Adding to the dry growth by cluster mergers, the compactions can increase the escape velocities to retain the SMBHs. The LRDs appear at $z \!\sim\! 8$, after the formation of FFB clusters, and disappear after $z \!\sim\! 4$ when the stellar mass is above $10^9 M_\odot$ by growing post-compaction blue disks around the nuclear LRDs. The LRD abundance is expected to be $\sim\! 10^{-5} \!-\! 10^{-4}\,{\rm Mpc}^{-3}$, increasing from $z \!\sim\! 4$ to $z\!\sim\! 8$.

astro-ph.GA↗

arXiv2026

When Irrelevant Text Matters: Affine Margin Shifts in Multimodal Large Language Models

Multimodal large language models (MLLMs) are frequently exposed to auxiliary textual context, the impact of which on visually grounded tasks remains underexplored. In this paper, we investigate the influence of task-irrelevant context by formulating it as a controlled intervention within a binary visual judgment framework. By maintaining an invariant prompt structure while varying auxiliary inputs, we observe that irrelevant text consistently biases model predictions across diverse benchmarks. To move beyond performance metrics, we characterize this sensitivity through a decision margin defined by the log-probability difference between binary candidates. Our analysis reveals a robust geometric regularity: contextconditioned margins follow a consistent affine transformation of their context-free counterparts. This finding demonstrates that irrelevant context does not manifest as unstructured stochastic noise but as a estimable distortion of model preference. We further interpret the fitted affine parameters as metrics for visual commitment preservation and directional answer bias. These findings provide a margin-level diagnostic view of irrelevant-context effects in MLLMs and offer a basis for future studies on noisy-context robustness

cs.CL↗

arXiv2026

FinTrace: Holistic Trajectory-Level Evaluation of LLM Tool Calling for Long-Horizon Financial Tasks

Recent studies demonstrate that tool-calling capability enables large language models (LLMs) to interact with external environments for long-horizon financial tasks. While existing benchmarks have begun evaluating financial tool calling, they focus on limited scenarios and rely on call-level metrics that fail to capture trajectory-level reasoning quality. To address this gap, we introduce FinTrace, a benchmark comprising 800 expert-annotated trajectories spanning 34 real-world financial task categories across multiple difficulty levels. FinTrace employs a rubric-based evaluation protocol with nine metrics organized along four axes -- action correctness, execution efficiency, process quality, and output quality -- enabling fine-grained assessment of LLM tool-calling behavior. Our evaluation of 13 LLMs reveals that while frontier models achieve strong tool selection, all models struggle with information utilization and final answer quality, exposing a critical gap between invoking the right tools and reasoning effectively over their outputs. To move beyond diagnosis, we construct FinTrace-Training, the first trajectory-level preference dataset for financial tool-calling, containing 8,196 curated trajectories with tool-augmented contexts and preference pairs. We fine-tune Qwen-3-8B/32B using supervised fine-tuning followed by direct preference optimization (DPO) and show that training on FinTrace-Training consistently improves intermediate reasoning metrics, with DPO more effectively suppressing failure modes. However, end-to-end answer quality remains a bottleneck, indicating that trajectory-level improvements do not yet fully propagate to final output quality.

cs.AI↗

arXiv2026

AgentOPSD: Recursive Self-Distillation for Agentic Reinforcement Learning

Reinforcement learning (RL) with verifiable rewards constructs trajectory-level advantage estimates, yet it often fails to credit the few pivotal decisions that determine outcomes in long-horizon, multi-turn agentic tasks. Recent work introduces privileged self-distillation for credit assignment, providing denser supervision, but it remains unclear how such local signals should represent sequential credit. We propose AgentOPSD, a critic-free, recursive method for turn-level credit assignment in agentic reinforcement learning. AgentOPSD aggregates token-level teacher-student log-probability gaps into turn-level evidence and recursively updates a Bayesian belief state in log-odds space. This yields a principled reweighting scheme that converts sparse outcome supervision into turn-level credit signals and identifies pivotal turns through the marginal belief revision between consecutive states. The method is fully compatible with standard policy optimization and requires neither an additional critic nor extra rollouts. We evaluate AgentOPSD on ALFWorld, WebShop, and Search-QA using Qwen2.5 models at two scales (3B and 7B). AgentOPSD outperforms GRPO and strong self-distillation baselines, achieving 89.1% success on ALFWorld with Qwen2.5-7B. Ablation studies attribute the gains to turn-level aggregation and history-dependent recursive belief updates.

cs.AI↗

arXiv2026

Distill Where You Fail: Recovering Learning Signals of Negative RL-Groups from Adaptive Teacher Guidance

Reinforcement learning with verifiable rewards (RLVR) has become a standard paradigm for post-training large language models (LLMs). While Group Relative Policy Optimization (GRPO) is widely adopted, it suffers from sparse reward signals and loses gradients entirely when all responses within a group receive identical rewards. On-policy distillation (OPD) offers a natural remedy by providing dense, token-level supervision from a teacher model. However, naively combining GRPO with OPD leads to degraded performance, due to three underlying causes: not all samples benefit from distillation; fitting too quickly to the teacher undermines the exploratory capacity of RL; and OPD's advantages are asymmetric, suppressing most tokens. To address these challenges, we propose RSTG (Recovering Learning Signals via Adaptive Teacher Guidance), which applies distillation selectively and precisely where it matters most. At the sample level, OPD is restricted to negative zero-variance prompts with each sample weighted by the teacher's confidence score. At the token level, distillation targets only tokens with high student entropy or large teacher-student divergence. We further augment training with SFT on correct trajectories generated by the teacher model, injecting positive gradient signals where RL yields none. Experiments demonstrate that RSTG substantially outperforms naive GRPO+OPD by +4.02% on math and +3.05% on code.

cs.CL↗

arXiv2026

VAD: Attributing Visual Evidence for Target Reconstruction in Multimodal On-Policy Distillation

Multimodal on-policy distillation (OPD) transfers fine-grained visual knowledge by supervising student-generated trajectories with a privileged-view teacher. Yet its next-token corrections are source-mixed, combining visual signals with linguistic priors and teacher-specific effects. The key challenge is to estimate which corrections are supported by visual evidence, not merely where or how strongly to distill. We introduce Visual Attribution Distillation (VAD), a counterfactual target-reconstruction algorithm that estimates the visually attributable part of a teacher correction. At each student-generated prefix, VAD evaluates the same fixed teacher with the relevant evidence present and removed. The corresponding change in centered log-probabilities defines ut, a signed proxy for the visual evidence direction that estimates how revealing the evidence supports or refutes candidate tokens. VAD projects the original correction onto this proxy to obtain an intervention-aligned component and a proxy-unexplained residual, then reconstructs a student-anchored target from the former. During training, this reconstructed target supplies the primary supervision signal, while the privileged teacher contributes a weak regularizer. Across six fine-grained visual benchmarks at 4B and 9B scales, VAD outperforms direct privileged-view distillation and visual-advantage weighting. Token- level and controlled-target analyses show that the proxy-aligned component is enriched in task-relevant visual corrections and yields stronger target shifts, especially when evidence refutes a mistaken answer. These results support counterfactual target reconstruction as an effective alternative to source-mixed supervision.

cs.CV↗

arXiv2026

SkillRise: Agentic Reinforcement Learning for Cross-Task Skill Evolution

Large language model agents often encounter related yet distinct tasks that share reusable solution patterns. Yet standard agentic reinforcement learning treats tasks as independent episodes, while existing approaches to skill learning either focus on repeated attempts of one task or use pipelines with multiple stages that entangle extraction, retrieval, and execution. We introduce SkillRise, a unified reinforcement learning framework for learning skills across tasks. SkillRise organizes related instances into progressively challenging sequences and uses a single policy to alternate between task solving and curating an evolving skill document passed directly to the next task. Decoupled credit assignment across tasks supervises solving with the current task outcome and curation with discounted downstream outcomes. Experiments on ALFWorld, WebShop, and ScienceWorld show that SkillRise achieves the strongest Pass@1 performance among the compared methods, with gains over the strongest baseline ranging from 2.3 to 8.5 percentage points. Although trained across distinct tasks, its learned curation policy remains effective for repeated attempts on the same task. Further analysis reveals scaling at test time across tasks: performance improves with longer sequences of related tasks even when each task is attempted only once. This trend suggests that SkillRise reuses transferable skills across tasks rather than benefiting from repeated sampling of the same task. SkillRise further retains strong performance while substantially reducing the runtime overhead of skill learning pipelines with multiple stages. Together, these results provide a simple and efficient training paradigm for LLM agents to extract, refine, and reuse transferable skills across tasks.

cs.LG↗