arXiv subjects
Publications and source records attributed to Siyu Chen.
We experimentally demonstrate simultaneous coherent transmission across 1-$μ$m and C+L bands, carrying $54\times125$-GBaud PDM-16QAM C+L channels and $2\times15$-GBaud PDM-QPSK 1-$μ$m channels over 50 km, highlighting 1 $μ$m as a promising additional transmission window while avoiding gas-absorption constraints in C/L-band of HCFs.
Looped transformers reuse a shared block of layers to increase computational depth without increasing the number of trainable parameters. They are trained with backpropagation through time (BPTT). To save training compute, BPTT is often truncated to the last few loops, but truncation yields biased gradients that may weaken the benefit of looping. In this work, we study the value of truncation in looped transformer training in a systematic way: we jointly vary model size, training tokens, and forward and backward loop counts across nearly 600 training runs and fit a unified scaling law that captures the effects of looping and truncation through effective parameters. Despite the training-compute savings of truncation, our fitted law favors full BPTT when forward and backward loop counts are jointly optimized. In matched-compute runs, full BPTT through fewer loops consistently outperforms truncated settings of the same compute. The law further shows that at fixed model size, the compute-optimal loop count and training-token budget both grow with compute. When model size is also free, the law identifies two fully trained loops as the best looped setting. Experimentally, this setting achieves lower validation loss than the compute-optimal non-looped model with about 27% fewer non-embedding parameters and only 8% more training compute.
Efficient and accurate evaluation of electron-phonon interactions is essential for predictive superconductivity calculations, but dense Brillouin zone sampling requirements pose a major computational bottleneck. To overcome this, we introduce an accelerated, highly scalable framework that combines a double delta truncation scheme with non-uniform Farey grids in both $\mathbf{k}$ and $\mathbf{q}$ space. The truncation strategy significantly reduces computational cost by restricting integration to states satisfying energy constraints near the Fermi level, while the Farey grids dynamically resolve sampling errors from localized scattering and stabilize convergence. Applied to elemental aluminium and several high-pressure hydrides, this combined methodology reduces computational cost by up to two orders of magnitude and yields electron-phonon coupling constants and superconducting critical temperatures Tc typically within about 5% of full-grid benchmarks. This approach overcomes sampling bottlenecks associated with dense uniform grids, making accurate, low-cost predictions of superconductivity in complex quantum materials feasible.
Multimodal reinforcement learning with verifiable rewards (RLVR) provides reliable task-level supervision, yet offers limited guidance for localizing errors in visual reasoning. On-policy self-distillation (OPSD) complements this supervision with dense token-level corrections from a privileged teacher equipped with an evidence-centered visual view. However, the resulting corrections may conflate visual evidence with non-visual shifts induced by linguistic priors, response style, or the privileged view itself. We formalize this phenomenon as privileged-correction contamination: a token-level correction is contaminated when its update direction cannot be consistently attributed to the privileged visual evidence. Directly distilling contaminated corrections may therefore reinforce erroneous token-level updates. To address this issue, we propose reinforcement learning with visually attributable on-policy self-distillation, an evidence-consistent paradigm that augments task-level RLVR supervision with privileged token-level corrections grounded in visual evidence. We instantiate this paradigm with Visual-Attribution Gating (VAG), which uses the verifier-derived advantage as the optimization anchor and validates each privileged correction against the contrast between evidence-present and evidence-degraded teacher views. By admitting only corrections directionally supported by this visual contrast, VAG preserves visually grounded guidance while suppressing inconsistent residuals. Experiments across six visual reasoning benchmarks show that VAG consistently outperforms GRPO and OPSD methods, establishing new state-of-the-art average performance.
Retrieval-augmented generation (RAG) trained with reinforcement learning (RL) is a promising way to ground large language models in medical evidence for clinical diagnosis. However, existing RAG-RL methods face two problems: (1) their process reward is largely singleview, as most methods inherited from opendomain multi-hop QA reward only the quality of the retrieved information; (2) assigning a reliable process reward is hard, as hard matching gives zero to clinically correct but non-verbatim steps such as "Allergic Sinusitis" versus "Allergic Rhinitis", and the alternative of strong-teacher guiding is bounded by the teacher's own medical competence. To address these issues, we propose C-MIG, a Multiview Information Gain-based framework for Clinical diagnosis. We identify two complementary and crucial intermediate steps in clinical diagnosis: what to retrieve and how to refine. We reward both by the information gain each brings, measured as the increase in the predicted likelihood of the ground-truth answer under a base, frozen reference model. The frozen scorer yields a stable and reproducible reward that needs no extra reward model or teacher. Experiments on four medical benchmarks show that C-MIG outperforms RAGRL baselines in and out of domain and stays competitive with far larger medical-specialist and general-purpose LLMs, at only about 3.4% training overhead and no inference-time cost.
Reconstructing reference videos as editable dynamic 3D scenes is an important path toward controllable video generation for filmmaking: these scenes provide modifiable, renderable visual proxies that enable explicit control over scene layout, subject actions, and camera behavior. To evaluate agents' ability to reconstruct editable dynamic 3D scenes from dynamic reference videos, we introduce CineRecon, a benchmark comprising 50 reference videos and 360 annotated evaluation clips. The reference videos span film, real-world footage, and games. Given a complete video, an agent uses tools to construct, animate, render, and refine its reconstruction. We evaluate this task under three representation settings: Proxy, Simplified, and Detailed. These settings increase geometric and appearance requirements while retaining the same requirements for spatial relations, camera behavior, and subject dynamics. This design tests whether agents can reproduce the events in a video when detailed appearance is unnecessary, and how performance changes as representation requirements increase. For evaluation, we construct checklists that associate observable properties and events with specific entities and times. A vision-language judge compares timestamp-aligned reference and rendered frames across 11 categories, grouped into camera, static, and dynamic consistency dimensions. We evaluate 12 models using their corresponding agent harnesses and find failures across all three dimensions. Subject motions and interactions remain difficult to reproduce even under the Proxy setting.
We present Pumpire, a unified benchmark for evaluating metric point-pair distance estimation capability of both image- and video-level 3D foundation models, with or without depth priors. In contrast to previous approaches that normally evaluate depth and camera intrinsics separately or evaluate point-clouds with geometric similarity metrics, which cannot directly reflect models' point-to-point distance estimation capability, Pumpire directly assesses point-to-point distances from the reconstructed geometry. To this end, we collect a large-scale and diverse dataset (pumpire-6k) comprising 100 real-world scenes, each annotated with physically measured point-pair distances and containing 64 frames, for a total of 6,400 frames. Building on this dataset, we establish a holistic evaluation protocol that covers both image- and video-level 3D foundation models and enables direct assessment of point-pair distance errors and cross-setting comparison. We conduct extensive experiments across 29 baseline configurations of representative 3D foundation models and provide a comprehensive analysis of the results. By offering this benchmark, we target the more fundamental ability to perceive and estimate physical scale in the reconstructed 3D space, which prior evaluation protocols have largely overlooked. The project page can be found at https://pumpire.github.io/
Visual imitation policies must learn which image regions matter for action, whether applying the same operation to different objects or selecting a language-specified target. We introduce ResCue (Residual Spatial Cueing), which converts language-derived target information into a spatial condition for visual policy learning. Pretrained grounding produces a target mask; an independent cue encoder adds its features to spatial RGB features before further visual processing and pooling. This retains scene context while making the target explicit. With Diffusion Policy (DP), ResCue achieves 56.5% average success across 15 simulated tasks, versus 33.1% for DP and 40.% for the strongest baseline, $π_0$, with gains in both ordinary manipulation and instruction-dependent object selection. On three real-world bimanual tasks, it reaches 60.0%, compared with 31.7% for $π_0$. Matched-mask ablations favor residual fusion over input-level and pooled alternatives, while test-time interventions show that target-correct guidance matters for execution.
Self-evolution offers a scalable path to stronger reasoning: a pretrained language model improves itself with only minimal external supervision. Yet existing methods either depend on extensively curated or teacher-generated training data, or, when the generator runs unsupervised, reward it by a difficulty heuristic that need not improve the solver. We introduce INFUSER, an iterative co-training framework with two co-evolving roles: a Generator that drafts questions and reference golden answers from a pool of unstructured, automatically collected documents, and a Solver that improves by training on them. The solver is trained with standard correctness rewards against the generator-provided answers, while the generator is rewarded by an optimizer-aware influence score that measures whether each proposed question would actually improve the solver on the target distribution. Because this continuous, noisy influence score is poorly served by standard GRPO, we propose DuGRPO, a dual-normalized variant of GRPO, for generator training. Together, these turn the document pool into an adaptive curriculum that favors questions useful to the current solver, not just hard ones. On Qwen3-8B-Base, INFUSER outperforms strong self-evolution baselines with over 20% relative improvement on Olympiad and SuperGPQA benchmarks, and an 8B INFUSER co-evolving generator outperforms a frozen 32B thinking generator on math and coding. Ablations confirm each design choice is necessary, and two extensions, applying INFUSER to an instruction-finetuned anchor and augmenting it with rule-verifiable RLVR data, further demonstrate the flexibility and generalizability of the framework. Code is available at https://github.com/FFishy-git/INFUSER.
The empirical success of pretraining large language models has inspired a deeper investigation into the underlying loss landscapes and the optimization dynamics. Recent empirical and theoretical study suggest that the training loss landscape often exhibits a "river-valley" structure, which features a low-loss manifold (river) flanked by sharp orthogonal directions with higher loss (mountains). In the long term, the optimization progress is determined primarily by the progress along the river. Within such a landscape, gradient descent with large learning rates can move faster along the river despite high apparent loss due to vertical oscillations, while a subsequent sharp decay in the learning rate suppresses these oscillations, revealing genuine optimization progress. This explains the recent success of warmup-stable-decay (WSD) learning rate scheduler which, unlike cosine scheduling, keeps stable high learning rate and decays before producing intermediate checkpoints. Building on this foundation, in this work we take a step further and study the role of momentum within such a loss landscape. We establish theoretical analysis that characterizes how momentum accelerates optimization by stabilizing large learning rates that can not be tolerated by vanilla GD without deviating significantly from the river. The enabled large learning rate in-turn gives greater speed along the river and makes faster essential progress in the long run. Another intriguing observation from theory is that for a river-valley landscape with very flat and slow-spinning river, the momentum itself does not contribute directly to acceleration in terms of the speed of tracking the river, while the main acceleration comes from the admissible larger learning rate.
Multi-turn jailbreak attacks demonstrate that harmful intent can be distributed across dialogue, yet existing methods obscure what conversational mechanisms drive vulnerability. We introduce BLUEPRINT, a safety-evaluation framework separating a factorized social-influence strategy space from WORLDVIEWSIM, a cross-turn situational context module. Monte Carlo Tree Search optimizes turn-level combinations of 18 theory-grounded influence factors across a four-turn trajectory. Across six frontier models, BLUEPRINT achieves near-ceiling ASR on major open-weight and proprietary models, while requiring the fewest average queries (2.46). The resulting trajectories further reveal model-specific vulnerability among resistant targets: each responds to distinct influence factors and strategy transitions, yet all share a common recovery pathway-shifting toward concrete, executable task framing consistently escapes hard-refusal states. Ablations confirm operational cues matter most: making requests actionable has the largest impact, gain framing is unusually potent, and some legitimacy appeals can backfire. These findings suggest robust multi-turn safety requires monitoring not only harmful content, but also how dialogue state makes unsafe requests appear concrete and locally executable.
Accurate large-scale wildfire spread modelling requires models that capture both the environmental conditions associated with fire occurrence and the local dynamics of fire propagation. We propose a three-stage framework that combines a Random Forest (RF) model with a cellular automaton (CA). First, an RF model trained on the 2021 Canadian fire season estimates daily pixel-level fire-occurrence probabilities. Second, quantile gradient boosting models provide optional spread-rate priors for sensitivity analysis. Third, an RF-informed CA combines the RF probability layer with neighbourhood-driven spread on a 5 km grid. The RF model achieved AUC values of 0.725--0.795 on the 2022--2024 datasets, while the RF-informed CA achieved substantially higher spatial overlap than the evaluated CA-only baselines in the 2023 simulation. A higher-resolution simulation provides an additional qualitative assessment of local spatial errors. These results suggest that combining RF-derived probabilities with local CA spread can improve large-scale wildfire simulations under the tested conditions.
Task-specific LoRA adapters offer a modular way to specialize large language and vision-language models. However, existing adapter composition methods are mostly static and cannot adapt to individual test inputs. To address these issues, we propose \textbf{LiST}, a label-free test-time LoRA fusion framework that converts an existing LoRA bank into a target-conditioned local simplex and searches sample-specific fusion weights at inference time. LiST builds joint task representations from LoRA parameter anchors and prompt-level behavior vectors, retrieves neighboring adapters as a local search space, and performs branch-preserving fusion without updating the backbone or adapters. Candidate weights are selected by a prompt-level energy with prior, geometric, and stochastic-consistency constraints, and are deployed only when they pass a safe acceptance rule. Otherwise, LiST falls back to a target-conditioned prior. Experiments on multimodal and language benchmarks show that LiST outperforms static LoRA merging and conventional test-time adaptation baselines, while preserving task-specific adapter utility and improving robustness on unseen tasks.
In this work, we perform a systematic, time-dependent study of the gamma-ray emission from four representative middle-aged SNRs (W51C, IC~443, W44, W28), incorporating both CRs within the remnant shells and escaped CRs interacting with surrounding molecular clouds. We compare our results with GeV--TeV gamma-ray observations from Fermi-LAT, H.E.S.S., MAGIC, and LHAASO, including a dedicated analysis of the Fermi-LAT data for regions A and B associated with W28. We find that the observed spectra favor steeper CR injection spectra with indices of \(α\sim4.2\)--\(4.3\), maximum proton energies of $\sim$ \(100\)--\(300\) TeV, diffusion coefficients below the Galactic average, and CR acceleration efficiencies from a few to tens of percent. In particular, the VHE emission detected by LHAASO from W51C is more naturally explained by escaped CRs interacting with a nearby molecular cloud. We also investigate the contribution of escaped CRs to the VHE emission from IC~443, W44, and W28. We further demonstrate that escaped CRs can substantially enhance the TeV neutrino flux from middle-aged SNRs, improving their prospects as potential neutrino sources. These results provide new constraints on CR acceleration and escape in middle-aged SNRs and highlight the important role of escaped CRs in shaping their high-energy gamma-ray and neutrino emission.
Achieving virtual agents capable of automating tasks across diverse digital environments remains a pivotal challenge in Embodied AI. While Multimodal Large Language Models (MLLMs) offer enhanced visual perception and reasoning, their agentic deployment faces three challenges: costly data annotation, imprecise action-intent alignment, and inefficient exploration from discarded failed trajectories. To address these, we introduce Iron, an intent-aligned, self-improved, and annotation-efficient framework for training GUI agents. Iron employs a novel dual learning strategy that utilizes a stepwise cycle-consistent (SCC) reward to achieve fine-grained alignment between low-level actions and high-level intents, thereby improving instruction grounding and intent understanding. Concurrently, Iron introduces a hindsight reproduction mechanism to repurpose failed trajectories for training, improving both learning efficiency and task diversity. Extensive experiments demonstrate that Iron-trained generalist agents consistently improve performance on cross-environment and cross-device tasks, outperforming models trained with three times more data. Iron also achieves a substantial 25.06% relative improvement on unseen web tasks, with further gains observed on inherently complex tasks, demonstrating the feasibility of building more capable virtual agents.
Personalized interpretation of medical reports has emerged as an increasingly important need among patients. Addressing this need requires both evidence-grounded medical factuality and context-dependent patient communication, yet existing medical vision-language tasks do not adequately capture these dual requirements. To bridge this gap, we introduce Patient-oriented Medical Report Interpretation (PMRI), a novel open-ended multimodal generation task that requires models to explain medical reports in accurate and accessible language based on a user's query and dialogue history. These two objectives differ fundamentally in their verifiability, yet remain tightly coupled, making them difficult to optimize jointly under conventional supervised fine-tuning and holistic reinforcement learning paradigms. To address this challenge, we propose G-CARL, a grounded, checklist-aligned reinforcement learning framework that combines multi-source retrieval for atomic claim verification with context-aware, instance-specific weighted checklists for response coverage, providing structured supervision for factuality, user-demand satisfaction, and expression quality without constraining response diversity. We further construct MMedReport, a real-world PMRI benchmark, along with a clinician-designed three-dimensional evaluation protocol. Extensive experiments demonstrate that G-CARL consistently outperforms existing post-training baselines in overall quality, claim-level precision, and checklist recall. Pairwise preference evaluation by clinicians further confirms that G-CARL produces interpretations that are more accurate and better aligned with patient needs.
The remarkable optoelectronic properties of metal halide perovskites are closely linked to their unusually soft and polar chemical bonds that enable both strong electron-phonon interactions and ion migration. Yet these two defining characteristics have largely been treated as independent consequences of the same underlying chemical bonding. Here we show that they originate from a common electronic-structure framework by developing a general description linking lattice dynamics, electron-phonon coupling, and halide ion migration across representative Pb-based, Sn-based, and double perovskites. Spectrally resolved phonon-mode contributions demonstrate that the low-frequency shearing modes dominate halide migration, whereas high-frequency stretching modes govern carrier scattering through the Fröhlich interaction in all three compositions. We introduce an orbital hybridization descriptor to unify these findings, which connects metal-halide bonding characteristics with the migration barrier energies and Fröhlich coupling strengths, indicating a cooperative evolution of these two properties. These findings provide a generalized microscopic mechanism for simultaneously optimizing charge and ionic transport in soft semiconductors.
Probabilistic extensions of Datalog enable static analyses such as pointer analysis, data race detection, and side-channel analysis to rank alarms by likelihood, but this added expressiveness also introduces a new challenge absent from deterministic analyses: the final output may contain alarms that are individually plausible yet mutually inconsistent, because marginal probabilities do not guarantee joint satisfiability. As a result, developers may spend effort investigating combinations of alarms that can never co-occur in any possible world. We address this problem by formalizing such inconsistencies as minimal unsatisfiable subsets (MUSes) and introducing PPProbe, a conflict extractor specialized for probabilistic Datalog analyses. Rather than improving MUS enumeration in general, PPProbe exploits the structure of Datalog derivation graphs to guide the search toward likely conflicts and prune the search space through bottom-up UNSAT inference. We evaluate PPProbe on 70 benchmarks from power side-channel analysis, data race detection, semantic diffing, and Bayesian-network inference. The results show that PPProbe achieves 2.5 to 24 times higher throughput than state-of-the-art MUS enumerators, and that the conflicts it identifies yield a conservative estimate of false-positive reduction, filtering out an average of 47.7% of mutually inconsistent alarms.