Search arXiv⌕ Search

arXiv subjects

Zeyu Chen

Publications and source records attributed to Zeyu Chen.

At least 19 recordsLinked to original sources
arXiv2026

Every Token Needs Its Own Teacher: Multi-teacher Adaptive On-policy Self-distillation

On-policy Self-distillation (OPSD) has proven effective for post-training large language models by constructing a teacher conditioned on privileged references. However, existing approaches keep the teacher fixed either throughout training by consistently exposing it to the full reference, or within each response by adaptively varying partial reference exposure only across responses. In this paper, we argue that neither of them is sufficient. Instead, every token needs its own teacher, as the teacher preference can vary token by token within the same response. Through a teacher-intervention pilot study, we validate that the teacher preference differs across tokens, and reveal that a preferred teacher can be identified for each token through majority voting and KL-based selection. Motivated by these findings, we propose Multi-teacher Adaptive On-policy Self-Distillation (MTA-OPSD), the first token-level adaptive OPSD method. Our method constructs a multi-teacher pool consisting of candidate teachers, and enables every token to adaptively select its preferred teacher during training. Experiments with Qwen3 models on mathematical reasoning tasks show that MTA-OPSD consistently outperforms existing OPSD approaches with an average gain of up to $+5.9$ over the base model, highlighting the importance of adapting teacher supervision at the token level. The code is available at https://github.com/xingzhejun/MTA-OPSD.

cs.CL↗
arXiv2026

Observational Evidence for Anisotropic Metal Excess around Galaxies

The exchange of matter and energy between galaxies and their surroundings drives the cosmic baryon cycle, yet mapping metal transport remains an observational challenge. While simulations predict that galactic winds escape anisotropically along minor axes, evidence for chemical enrichment in neighboring galaxies is limited. We analyze 1,433 galaxy pairs from the Dark Energy Spectroscopic Instrument survey and detect a gas-phase metallicity excess of 14.6% $\pm$ 3.7% to 24.2% $\pm$ 2.6% in neighbors aligned with the minor axis of massive primary at projected separations of 15--60 kpc, qualitatively consistent with IllustrisTNG simulation. Permutation tests indicate that this excess varies from a marginal detection ($>$92% confidence) at closer separations (15--30 kpc) to a significant signal ($>$98% confidence) at intermediate separations (30--60 kpc). Here, we show that these results provide empirical constraints on galactic feedback models and are consistent with a scenario where anisotropic outflows can enrich nearby galaxies, a mechanism that complements other environmental processes.

astro-ph.GA↗
arXiv2026

Deconstructing Off-Policy Ratios: Entropy-Normalized Trust Regions for Asynchronous Reinforcement Learning

Asynchronous reinforcement learning (RL) accelerates large language model (LLM) post-training by overlapping rollout generation with policy optimization, but the resulting stale, off-policy data destabilizes optimization and can cause policy collapse. Existing methods gate tokens by ratio magnitude alone, applying one threshold at every position. We show that the ratio's natural scale is set by token entropy, so deviations from mid-trajectory weight updates stay within this scale and carry genuine exploration. We further identify an overlooked low-entropy regime that breaks this scaling, where a near-zero probability amplifies train--inference mismatch into noise far beyond what the local entropy admits. A magnitude threshold admits this noise and discards the exploration. We therefore propose the Entropy-Normalized Trust Region (ENTR). Across long-horizon agentic tasks and mathematical reasoning benchmarks, ENTR outperforms existing asynchronous methods. It improves avg@1 on BrowseComp-Plus by $6.9\%$ over the strongest baseline, trains stably up to $30$ policy versions of staleness, and matches synchronous GRPO at a $2.6\times$ speedup.

cs.AI↗
arXiv2026

Fountain pattern of baryon cycle revealed in galaxy ecosystems

Baryons in galaxy ecosystems are believed to undergo continuous cycles of inflow and outflow, forming fountain-like patterns that encode key information about how galaxies acquire matter from their environments and respond through feedback. The presence of such baryon cycles has been inferred from pieces of observational evidence, but a concrete understanding remains elusive because individual galaxy ecosystems are diverse and dynamic. Here we introduce a stacking method that combines baryonic fields across ensembles of individual galaxy ecosystems to suppress irregularities and reveal the underlying pattern. Applied to a cosmological hydrodynamic simulation, this approach unveils strikingly regular patterns in gas properties across the full spatial extent of galaxy ecosystems, in close agreement with those inferred from observations. This method is straightforward to implement, allowing the processes shaping the gas-cycling pattern to be fully understood within the structure-formation paradigm, and a solid framework to be constructed for linking simulated galaxy ecosystems with observations.

astro-ph.GA↗
arXiv2026

BRACE: Anchored Bellman-Residual Correction for Stale Critics in Asynchronous RL

Asynchronous reinforcement learning has become the standard way to scale training for large language models (LLM), but the resulting policy lag biases the critic toward the stale behavior policy. Existing work on asynchronous LLM training corrects the actor and leaves this bias unaddressed, while the off-policy value correction of classical RL does not carry over to long-horizon agentic tasks, since a short correction horizon leaves the regression target free of the reward and a long one lets the product of importance ratios drift exponentially with the trajectory length. We propose BRACE, an anchored Bellman-residual correction for stale value models. BRACE bounds the correction horizon to a prefix of policy tokens and anchors a constant-weight Monte-Carlo tail beyond it, which separates policy correction from reward propagation. BRACE delivers a $9.8\%$ relative improvement in mean@1 on BrowseComp-Plus over the strongest baseline, runs $2.46\times$ faster per step than synchronous training, and remains stable $50$ updates off-policy.

cs.LG↗
arXiv2026

Valley3: Scaling Omni Foundation Models for E-commerce

In this work, we present Valley3, an omni multimodal large language model (MLLM) developed for diverse global e-commerce tasks, with unified understanding and reasoning capabilities across text, images, video, and audio. A key feature of Valley3 is its native multilingual audio capability for e-commerce, developed by extending vision-language models to better support crucial audio-visual tasks, particularly in short-video scenarios. To achieve this, we carefully design a four-stage omni e-commerce continued pre-training pipeline, through which Valley3 progressively acquires audio understanding, cross-modal instruction-following, e-commerce domain knowledge, and long-context reasoning capabilities, ultimately evolving into an omni model for diverse e-commerce scenarios. Then, we further improve Valley3 through post-training to encourage long-chain reasoning with controllable reasoning modes, enabling one non-thinking mode and three distinct levels of thinking, thereby balancing inference efficiency in simple scenarios with deep reasoning for complex applications. Moreover, we equip Valley3 with agentic search capabilities to proactively invoke search tools and acquire task-relevant information for e-commerce deep research tasks. To comprehensively assess the capabilities of Valley3, we construct an omni e-commerce benchmark spanning 6 tasks. Experimental results show that Valley3 consistently outperforms strong baselines on our in-house and open-source e-commerce benchmarks, while remaining competitive on general-domain benchmarks. Our code and model weights are available at https://github.com/bytedance/Valley.

cs.AI↗
arXiv2026

Symmetry Discovery in Quantum Learning: Observable-Level and Task-Level Inference from Finite Measurements

Symmetry reduces the capacity of a quantum learning model, but the imposed group must match both the measured information and the label transformation. We establish a finite-measurement theory for inferring this group from candidate transformations. The central structural result identifies observable-invisible transformations with the stabilizer of a projected state whenever the probe span is invariant. It turns recovered generators into a valid subgroup and identifies the continuous invisible space with its Lie algebra. For finite dictionaries, an unbiased shadow statistic distinguishes zero from positive squared expectation discrepancies with an inverse-gap measurement rate, improving the inverse-square-gap rate of uniform discrepancy estimation. A commuting qubit lower bound proves the gap dependence optimal at fixed snapshot scale, and simultaneous intervals support data-dependent tolerances. Task validation then tests either the joint distribution through a characteristic kernel or its encoded mean through a classical--quantum discrepancy. An exact group-average identity relates the latter to joint-state asymmetry and specifies its conversion to binary task breaking mass. Projection bias quantifies the cost of excessive symmetry, while an $\ell_1$ readout bound quantifies the capacity gained by relaxing it. At an invariant pure-state backbone, retained and nontrivial breaking sectors are Fisher-orthogonal. Ising-chain calculations connect finite-shot recovery, label-dependent symmetry, and physical sector drift. These results determine which symmetry the measurements support and provide the statistical and geometric basis for a subsequent release decision.

quant-ph↗
arXiv2026

Quantum Natural Gradient on Quotient Spaces

A parametrized quantum circuit reports its state geometry through a quantum Fisher information matrix (QFIM), often singular. A small Fisher value can reflect exact state-preserving redundancy, compression by the circuit chart, or weak intrinsic distinguishability, and these mechanisms call for different numerical treatments. We show that the circuit metric factors as $F=B^{*}MB$, where $B$ is the state-level circuit-to-orbit differential and $M$ is the intrinsic Fisher operator on the reachable orbit. The factorization identifies the exact kernel as $\ker B$, separates coordinate transfer from intrinsic geometry, and yields the condition for a circuit to realize an orbit-level quantum natural-gradient (QNG) direction. When the prescribed redundancy exhausts the Fisher kernel, the Moore--Penrose update is the minimum-norm horizontal lift of the quotient Riemannian gradient. At critical points with a locally diffeomorphic quotient-to-orbit map, chart singular values cancel from the linearized QNG operator while intrinsic anisotropy remains; in the trace-orthonormal full-control generator frame, excitation-gap anisotropy gives $κ_{\mathrm{QNG}}=κ_{\mathrm{Eucl}}$. Representation theory makes $M$ explicit on highest-weight, Slater, and fermionic-Gaussian orbits, and cominuscule fidelity flow becomes integrable, with conserved principal-defect ratios, cubic Lie-retracted convergence at $η=2$, and stability boundary $η=4$. Finite data impose a second boundary: an estimated QFIM and its confidence radius alone cannot distinguish an exact zero from a small physical mode, so the estimated spectrum alone cannot license hard projection. Under depolarization, inverse-Fisher scaling amplifies mean updates and fluctuations together and cannot restore update signal-to-noise. A redundant Slater/Givens circuit confirms exact transfer identities and illustrates finite-shot tradeoffs.

quant-ph↗
arXiv2026

Statistical Symmetry Release for Equivariant Quantum Learning

Hard symmetry constraints reduce model complexity, but can also erase label information. Statistical symmetry release determines when finite data and quantum measurements justify relaxing such a constraint, which directions to open, and how far to move. We connect global signal detection to local, loss-dependent improvement. A two-copy twirl--swap gate estimates task information in the symmetry-breaking complement with a dimension-independent copy count under paired-state and group-unitary access; reweighting the same records resolves representation sectors. An exact duality distinguishes this Hilbert--Schmidt signal from the larger signal accessible to bounded-outcome readouts. Local improvement is governed by the release gradient and a loss-corrected double-commutator matrix. Simultaneous confidence bounds convert empirical direction selection into certified descent, using either shared Pauli measurements or scalar probes with state-independent truncation bounds. Gaussian testing lower bounds quantify the cost of searching over unknown directions in the calibrated local experiment. Independent validation controls adaptively generated models, and a fast squared-loss bound preserves the approximation--estimation rate of a nested release path. On an eight-qubit Ising model, shared measurements certify release with 6300 times fewer shots than the specified scalar estimator on the tested budget grids. Quotient quantum natural gradient then controls parameter redundancy during training. Together, these results turn symmetry relaxation into a statistically justified model-selection decision.

cs.LO↗
arXiv2026

Rendering-in-the-Loop: An Execution-Driven Agent for Interactive Web Development

Multimodal large language models have achieved remarkable progress in front-end web development, generating interactive webpages from multimodal references such as screenshots and interaction videos. However, existing work largely emphasizes visual metrics such as aesthetics and layout similarity, while overlooking the more critical validation of interactive functionality. We present RILA, an execution-driven agent that puts browser rendering in the loop, iteratively editing generated code from runtime interaction feedback. RILA introduces an Action Interaction Verification (AIV) module that replays the reference interaction trajectory on the generated webpage to collect grounded execution-aware observations, and an Execution-aware Rendering Score (ERS) that jointly measures interaction correctness and visual fidelity to guide iterative optimization. We further build an execution-verified data synthesis pipeline that produces diverse, high-quality training data, offering gains complementary to inference-time optimization. On IWR-Bench, RILA consistently improves both interaction and visual fidelity across foundation models. Notably, with our training pipeline, RILA lifts the compact Qwen3.5-9B backbone from 40.40% to 57.52%, surpassing far larger one-shot generators, including the 1T-parameter Kimi-K2.6 (55.61%) and the proprietary GPT-5.5 (55.74%).

cs.CV↗
arXiv2026

Behavioral Memory under Symmetry in One-Way Quantum Automata

Under compact symmetry, observable behavior reduces to an invariant operator algebra, but its dimension is not yet classical memory: some coordinates are dynamically frozen, some invisible to threshold tests, and some already classical. We develop an operator-algebraic theory that separates these effects through three filters. For one automaton, behavior is the Hilbert--Schmidt pairing between prefix-reachable states and suffix-observable effects, whose rank equals the real Hankel rank without controllability or observability assumptions. Maximizing this invariant over a symmetry-constrained dynamical class gives a structural capacity controlled by the symmetry commutant: its center stores isotypic populations frozen by reversible dynamics, its traceless multiplicity blocks carry movable noncommutative coordinates, dissipation removes the unary spectral loss inside those blocks, and covariant mobility releases relative populations subject to component conservation. Operational realization then determines which surviving coordinates force probabilistic states. For a fixed nontrivial invariant readout, full mobility gives an exact dichotomy in worst-case state cost: a commutative invariant algebra costs exactly its dimension, whereas a noncommutative multiplicity block raises the unrestricted cost by exactly one state. Thus noncommutativity has a one-state worst-case classical price. The known four-letter quadratic-plus-one law at trivial symmetry is the fully mobile endpoint of this principle. Schur--Weyl duality further shows that different preserved symmetries on the same tensor-power Hilbert space can change the worst memory scale from polynomial to exponential, while fixed-weight modules give an exact Catalan law at half filling, with structural capacity equal to the Catalan count minus its central-sector correction.

cs.FL↗
arXiv2026

Efficient quantum compression for identically prepared states with arbitrary dimensio

Identical preparation creates permutation symmetry that can be used for lossless quantum compression. For $n$ copies of an unknown $d$-dimensional pure state, the tensor-power input lies in the fully symmetric subspace, whose dimension is polynomial in $n$ for fixed local dimension. Schur--Weyl duality isolates this subspace coherently, allowing the fixed representation labels to be discarded while retaining all information needed for recovery. The resulting encoder and decoder are exact on the input family. Because tensor-power states span the symmetric subspace, the achieved memory dimension is also necessary for reversible coherent compression, so the scheme is space-optimal. The encoder is implemented recursively with Clebsch--Gordan transforms, and only the branches reached by symmetric inputs need to be reproduced. Standard efficient synthesis of these transforms yields polynomial-size circuits for fixed local dimension and arbitrary target accuracy. Thus identically prepared pure states in arbitrary dimension admit lossless, space-optimal compression with an efficient circuit implementation.

quant-ph↗
arXiv2026

The State Cost of Classical Simulation of One-Way General Quantum Finite Automata

Under strict cutpoints, probabilistic finite automata (PFAs) and one-way general quantum finite automata (1gQFAs) recognize the same stochastic languages, shifting the theoretical focus to the state cost required for a classical PFA to simulate a 1gQFA. For an $n$-state ($n\geq 2$) 1gQFA, the state cost upper bound of classical simulation was previously known to be $n^2+3$, while the lower bound remained an open problem widely conjectured to be quadratic. After establishing a well-defined notion of classical simulation, we improve the existing state cost upper bound from $n^2+3$ to $n^2+1$. Subsequently, we introduce the concepts of shattering and dynamic shattering, which are used to determine the memory required for a PFA to recognize a language. Using these techniques, we prove that the state cost lower bound of the simulation reaches $n^2+1$ over a four-letter alphabet. With the upper and lower bounds thus matching, the problem is fully resolved.

cs.FL↗
arXiv2026

ECHO: Prune To Act, Trace To Learn With Selective Turn Memory In Agentic RL

Long-horizon language agents must repeatedly interact with tools, accumulate evidence, and make decisions under bounded context windows. Context-management methods make such rollouts feasible by simplifying past interactions through deletion, folding, or memory editing. However, when useful history is collapsed into compressed states, the reconstructed context may no longer reveal which earlier observations support a successful final answer. This creates a mismatch between bounded-context acting and outcome-based reinforcement learning: the policy acts on reconstructed context, while the learner lacks source-level provenance for assigning credit to the evidence that mattered. We propose ECHO, a selective turn-memory framework for traceable context reconstruction in Agentic RL. ECHO compresses each completed environment turn into a compact source-indexed memory record, reconstructs bounded policy contexts by selecting useful records, and reuses the selected source indices to route positive outcome credit to the final trajectory segment, reused evidence turns, memory findings, and memory-selection actions. On BrowseComp-Plus, ECHO reaches 43.4% held-out accuracy, outperforming GRPO at 28.9% and the rolling-summary baseline SUPO at 36.1%, while using fewer turns and lower trajectory volume than SUPO. The trained policy also improves zero-shot generalization across multi-objective QA, code generation, and deep information-seeking benchmarks on both dense and MoE backbones.

cs.LG↗
arXiv2026

Group-Reflective Self-Distillation for Agentic Reinforcement Learning

Reinforcement learning with verifiable rewards (RLVR) is effective for training large language model agents. However, terminal rewards provide only coarse trajectory-level supervision, leaving successful behaviors, recurring mistakes, and incidental choices entangled in the same outcome signal. Existing agentic self-distillation methods enrich sparse supervision with natural-language skills, but skills retrieved externally or extracted from a single trajectory by stronger models may mismatch current experience, exceed the policy's capability, or remain path-specific. We propose Group-Reflective Self-Distillation (GRSD), which derives capability-aligned and outcome-discriminative guidance from the policy's own verified rollouts. For each prompt, the policy reflects on each verified trajectory in an on-policy group, and a stop-gradient snapshot contrasts the resulting reflections from successful and failed rollouts to construct group-level privileged guidance. Conditioned on this guidance, a self-teacher refines turn-level credit assignment by modulating outcome-based advantages while preserving the verifier-determined learning direction. Experiments across multiple agentic environments and model scales demonstrate that GRSD consistently outperforms competitive baselines and generalizes more effectively to unseen tasks.

cs.AI↗
arXiv2026

Open-Set Visual Text Forensics via Sparse-Constraint Rectified Flow

Rapidly evolving Generative AI enables sophisticated visual text manipulations that increasingly evade current forensic detectors. Existing discriminative models often overfit specific forgery patterns, limiting their generalization to unseen, open-set attacks. To address this challenge, we propose a generative detector that localizes tampering by estimating the local restoration cost required to align a query image with authentic visual-text statistics, rather than by learning forgery-specific decision boundaries. Specifically, we introduce Sparse-Constraint Rectified Flow (SC-RF), a detector-oriented adaptation of Flow Matching for spatially sparse anomaly localization. We further mitigate data scarcity via self-supervised Artifact Injection and preserve high-frequency forensic traces using a pixel-space Forensic-DiT. Extensive experiments on three benchmarks show that our method achieves state-of-the-art performance, surpassing the runner-up by 3.2 and 4.8 percentage points in F1 and IoU, respectively. In particular, the proposed detector demonstrates strong zero-shot performance on challenging unseen text editing patterns. We further provide an auxiliary stress-test analysis showing that local harmonization produced by our model can weaken the statistical cues relied upon by existing detectors, offering a complementary vulnerability-analysis perspective.

cs.CV↗
arXiv2026

ACPO: Asymmetric Credit Policy Optimization via Mode-Local Entropy Surrogate

Outcome-supervised reinforcement learning scales to verifiable reasoning tasks, but trajectory-level rewards assign the same outcome signal to all sampled tokens, overlooking their unequal contributions to the reasoning process. Entropy provides a natural indicator of the model's decision state, yet using it for token-level credit assignment presents two key challenges: long-tail probabilities in large vocabularies corrupt both entropy values and gradients, and uncertainty carries distinct semantics across positive- and non-positive-advantage trajectories. We propose Asymmetric Credit Policy Optimization (ACPO), which replaces global entropy with the complement of the top-token probability as a mode-local proxy. Guided by gradient analysis, ACPO incorporates mismatch routing and saturation correction to shape policy updates into the desired asymmetric form, emphasizing uncertain decisions on positive trajectories while penalizing confident regions on failed ones. Theoretically, ACPO locally preserves the advantage direction while bounding surrogate error. Experiments on mathematical and coding reasoning benchmarks, including AIME 2025 and HumanEval Pro, show that ACPO consistently outperforms both entropy-aware methods (e.g., 80/20, GTPO) and strong outcome-supervised RL baselines (e.g., DAPO, SAPO).

cs.LG↗