arXiv subjects

Fei Zhao

Publications and source records attributed to Fei Zhao.

At least 19 recordsLinked to original sources

arXiv2026

Verify Before You Distill: Prompt-Level Teacher Gating for On-Policy Distillation

On-policy distillation (OPD) accelerates post-training by providing dense token-level supervision from a frozen teacher on the student's own rollouts. Vanilla OPD applies this supervision uniformly across prompts, without checking whether the teacher is reliable for each prompt. Because reverse KL is mode-seeking, a confidently wrong teacher can induce a strong yet misleading update. Distributional proxies, such as entropy or teacher-student likelihood agreement, measure uncertainty or agreement but do not directly verify outcome correctness. We introduce Teacher-Gated On-Policy Distillation (TGOPD), built on the principle that teacher reliability should be verified at the prompt level before dense supervision is admitted. TGOPD estimates reliability from a small set of verifier-scored teacher probes and routes each prompt exclusively to dense OPD when the reliability check passes or to verifier-grounded GRPO otherwise. Across 4B and 35B students in mathematics, code, and instruction following, TGOPD outperforms Vanilla OPD in all six single-domain settings and achieves higher seven-benchmark averages at both scales under multi-domain training. By using otherwise-idle teacher capacity for reliability estimation, TGOPD also reduces teacher-side compute waste in asynchronous OPD, increasing teacher-node GPU utilization from 9.8% to 78.9% in the measured 4B single-domain run.

cs.LG↗

arXiv2026

D$^3$-MOPD: Adaptive Dynamic Domain ScheDuling for Efficient Multi-Teacher Distillation

Multi-teacher on-policy distillation (MOPD) distills several domain-expert teachers into a single student by minimizing per-domain reverse-KL divergence on the student's own rollouts. Existing approaches typically fix the per-domain data mixture before training, overlooking the fact that different domains converge at substantially different rates: some plateau early while others continue to improve throughout the training budget. A fixed mixture therefore wastes compute on fast-converging domains and undertrains slower-converging ones. To address this, we propose D$^3$-MOPD (Dynamic Domain ScheDuling for MOPD), a zero-overhead scheduler that repurposes the per-domain reverse-KL signal already produced during training to adapt the domain mixture online. Running asynchronously outside the training process, an off-process watcher periodically tracks each domain's KL trajectory, estimates remaining headroom and current improvement rate, and accordingly adjusts the domain sampling ratios without altering the core training loop. Our D$^3$-MOPD scales naturally to arbitrary numbers of domains, and the expected benefit grows as more domains introduce more diverse convergence patterns for the scheduler to exploit. On a Qwen3.6-35B-A3B student distilled from four domain-expert teachers, D$^3$-MOPD closes 97% of the average student-to-teacher performance gap, compared with 63% for vanilla MOPD, reaches the same peak performance with an approximately 3$\times$ reduction in rollout steps, and surpasses the specialist teachers on three of seven benchmarks.

cs.LG↗

arXiv2026

Beyond Decision Boundaries: Relational Geometry Attacks on Contrastive Embedding Manifolds

Contrastive learning and Siamese embedding models have become the foundation of modern verification systems, where decisions are governed not by discrete classification boundaries, but by relational geometry in embedding space. However, existing adversarial attacks remain fundamentally classification-centric, overlooking the vulnerability of relational geometry. In this paper, we introduce a geometry-aware adversarial attack framework that reformulates attacks on contrastive systems as manifold-level relational corruption. Instead of targeting individual predictions, the proposed framework systematically distorts similarity organization within the embedding manifold by pushing positive pairs apart while simultaneously pulling negative pairs closer, ultimately collapsing and inverting pairwise similarity structure. To enable scalable deployment, we shift iterative online optimization into an offline adversarial geometry deformation prior learning stage and train a lightweight feed-forward generator that learns generalized geometry deformation patterns from the victim model. Once trained, the generator produces adversarial perturbations through a single forward pass without requiring online gradient computation, enabling real-time online attacks against similarity-based verification systems. Experimental results across multiple verification architectures demonstrate substantial degradation of verification performance together with severe manifold-level relational corruption. On the Markmatch verification system, the proposed attack reduces accuracy from 95.4% to 38.6% while completely reversing the positive-negative similarity structure.

cs.AI↗

arXiv2026

Evolving Safety Landscape of Multi-modal Large Language Models: A Survey of Emerging Threats and Safeguards

Multi-modal large language models (MLLMs) integrate heterogeneous modalities through modality alignment and fusion, enabling stronger understanding and reasoning. However, this architectural shift reshapes the safety landscape of machine learning. Increased model complexity and cross-modal interactions give rise to novel threats, including compromised modality integration, modality misalignment, and fused safety risks, reflecting shifts in threat modeling beyond uni-modal assumptions. These shifts, in turn, impose new constraints on safety solutions not captured by existing frameworks rooted in uni-modal learning. Motivated by these challenges, this survey provides a systematic analysis of the evolving safety landscape of MLLMs. We first propose a multimodal grounded taxonomy of safety threats and analyze shifts in threat models, covering adversarial attacks, data poisoning, jailbreaks, and hallucinations. We then summarize updated safety assumptions and organize recent advances in MLLM safety strategies accordingly. Finally, we discuss open challenges and future directions to inform the development of more principled and scalable safety mechanisms for multimodal systems.

cs.LG↗

arXiv2026

Group-of-Latents: Perceptual Video Compression at Extreme Bitrates via Masked Latent Generative Modeling

Most existing video compression algorithms follow a paradigm of transformation and quantization, optimizing the trade-off between distortion and bitrate. However, extremely low-bitrate compression remains an underexplored frontier where perceptual quality optimization under severely constrained coding resources has not been adequately addressed. In this paper, we propose a unified generative framework that leverages pre-trained Diffusion Transformer (DiT) priors to achieve high perceptual quality at extremely low bitrates. We first introduce a flexible Group-of-Latents (GoL) strategy within the latent space of a causal tokenizer, explicitly partitioning the latent stream into intra $I$-latents and inter $P$-latents. The Deep Compression Module (I-DCM) then encodes key $I$-latents to preserve perceptual anchors with minimal overhead. Building upon these anchors, the DiT-based Unified Latent Denoising Module (U-LDM) refines intra-frame textures and synthesizes $P$-latents from noise, reconstructing temporal dynamics at zero additional bitrate cost. Extensive experiments demonstrate that our method uniquely operates in the extreme-low-bitrate regime (e.g., (<0.005) bpp), achieving state-of-the-art perceptual fidelity with rich spatial details and robust temporal consistency. The code will be made publicly available.

eess.IV↗

arXiv2026

The capability of CSST in characterizing planetary atmospheres. I. transmission spectroscopy of hot Jupiters

Transmission spectroscopy has become a primary tool for probing exoplanetary atmospheres, enabling constraints on their chemical compositions and providing limited information on their thermal properties. We assess the potential of the upcoming Chinese Space Station Telescope (CSST) for exoplanet atmospheric characterization through transmission spectroscopy. Theoretical spectra of hot gas planets are generated and used to simulate slitless spectroscopic observations with the CSST across the ultraviolet-to-near-infrared range. Atmospheric retrievals performed on the simulated data are compared with the input models to assess the robustness and accuracy of parameter determinations. We find that multi-band observations across three wavelength channels, each with two transits can place meaningful constraints on key atmospheric parameters. For multi-band observations that account for correlated (red) noise, future CSST observations are expected to achieve constraints that are comparable to, or in some cases slightly weaker than, those of the Hubble Space Telescope (HST), depending on the noise level and observing strategy. We conclude that CSST will provide unique and complementary constraints on the chemical compositions and physical properties of exoplanetary atmospheres, particularly for atomic species, metal-bearing molecules, and scattering processes accessible in the UV and optical, thereby complementing JWST's infrared sensitivity to molecular species.

astro-ph.EP↗

arXiv2026

The 3D Architecture of a pair of 6:1 Resonant Brown Dwarfs around the Naked-eye star $\nu$ Ophiuchi

We present a revisiting study of the brown dwarf pair orbiting the naked-eye ($V=3.3$) K-giant $\nu$~Ophiuchi, located only 44\,pc from our Solar system. By jointly analysing archival radial-velocity measurements together with astrometric data from \textit{Hipparcos} and the \textit{Gaia} second and third data releases, we determine the three-dimensional architecture of the system and robustly constrain the masses of both companions. We find brown dwarf masses of $m_{\mathrm{b}} = 24.2^{+6.4}_{-2.8}\,M_{\mathrm{J}}$ and $m_{\mathrm{c}} = 26.8^{+4.3}_{-2.9}\,M_{\mathrm{J}}$. The mathematical constraint, derived from the posterior distribution of the mutual inclination based on MCMC samples, yields a mutual inclination of $\psi_{\mathrm{bc}}=46^{+27}_{-24}\!\,^{\circ}$, while direct calculations based on the maximum a posteriori and posterior median orbital parameters yield values of $\sim$$10^{\circ}$ and $\sim$$20^{\circ}$, respectively. Resonance analysis indicates that the two companions can still be trapped in a 6:1 mean-motion resonance in the maximum a posteriori configuration. To place an upper limit for the mutual inclination, dynamical stability analysis over a 1~Myr timescale further constrains it to be no larger than $\sim$$15^{\circ}$. Systems hosting brown dwarf pairs are rare, yet they provide important constraints on theories of planetary formation and dynamical evolution. Current detections suggest that brown dwarf pairs preferentially reside at large separations from their host stars and are more common in less mature systems. This supports a star-like formation pathway via gravitational instability in disk.

astro-ph.EP↗

arXiv2026

Talk2DM: Enabling Natural Language Querying and Commonsense Reasoning for Vehicle-Road-Cloud Integrated Dynamic Maps with Large Language Models

Dynamic maps (DM) serve as the fundamental information infrastructure for vehicle-road-cloud (VRC) cooperative autonomous driving in China and Japan. By providing comprehensive traffic scene representations, DM overcome the limitations of standalone autonomous driving systems (ADS), such as physical occlusions. Although DM-enhanced ADS have been successfully deployed in real-world applications in Japan, existing DM systems still lack a natural-language-supported (NLS) human interface, which could substantially enhance human-DM interaction. To address this gap, this paper introduces VRCsim, a VRC cooperative perception (CP) simulation framework designed to generate streaming VRC-CP data. Based on VRCsim, we construct a question-answering data set, VRC-QA, focused on spatial querying and reasoning in mixed-traffic scenes. Building upon VRCsim and VRC-QA, we further propose Talk2DM, a plug-and-play module that extends VRC-DM systems with NLS querying and commonsense reasoning capabilities. Talk2DM is built upon a novel chain-of-prompt (CoP) mechanism that progressively integrates human-defined rules with the commonsense knowledge of large language models (LLMs). Experiments on VRC-QA show that Talk2DM can seamlessly switch across different LLMs while maintaining high NLS query accuracy, demonstrating strong generalization capability. Although larger models tend to achieve higher accuracy, they incur significant efficiency degradation. Our results reveal that Talk2DM, powered by Qwen3:8B, Gemma3:27B, and GPT-oss models, achieves over 93\% NLS query accuracy with an average response time of only 2-5 seconds, indicating strong practical potential.

cs.AI↗

arXiv2026

QP-OneModel: A Unified Generative LLM for Multi-Task Query Understanding in Xiaohongshu Search

Query Processing (QP) bridges user intent and content supply in large-scale Social Network Service (SNS) search engines. Traditional QP systems rely on pipelines of isolated discriminative models (e.g., BERT), suffering from limited semantic understanding and high maintenance overhead. While Large Language Models (LLMs) offer a potential solution, existing approaches often optimize sub-tasks in isolation, neglecting intrinsic semantic synergy and necessitating independent iterations. Moreover, standard generative methods often lack grounding in SNS scenarios, failing to bridge the gap between open-domain corpora and informal SNS linguistic patterns, while struggling to adhere to rigorous business definitions. We present QP-OneModel, a Unified Generative LLM for Multi-Task Query Understanding in the SNS domain. We reformulate heterogeneous sub-tasks into a unified sequence generation paradigm, adopting a progressive three-stage alignment strategy culminating in multi-reward Reinforcement Learning. Furthermore, QP-OneModel generates intent descriptions as a novel high-fidelity semantic signal, effectively augmenting downstream tasks such as query rewriting and ranking. Offline evaluations show QP-OneModel achieves a 7.35% overall gain over discriminative baselines, with significant F1 boosts in NER (+9.01%) and Term Weighting (+9.31%). It also exhibits superior generalization, surpassing a 32B model by 7.60% accuracy on unseen tasks. Fully deployed at Xiaohongshu, online A/B tests confirm its industrial value, optimizing retrieval relevance (DCG) by 0.21% and lifting user retention by 0.044%.

cs.IR↗

arXiv2026

Decouple Searching from Training: Scaling Data Mixing via Model Merging for Large Language Model Pre-training

Determining an effective data mixture is a key factor in Large Language Model (LLM) pre-training, where models must balance general competence with proficiency on hard tasks such as math and code. However, identifying an optimal mixture remains an open challenge, as existing approaches either rely on unreliable tiny-scale proxy experiments or require prohibitively expensive large-scale exploration. To address this, we propose Decouple Searching from Training Mix (DeMix), a novel framework that leverages model merging to predict optimal data ratios. Instead of training proxy models for every sampled mixture, DeMix trains component models on candidate datasets at scale and derives data mixture proxies via weighted model merging. This paradigm decouples search from training costs, enabling evaluation of unlimited sampled mixtures without extra training burden and thus facilitating better mixture discovery through more search trials. Extensive experiments demonstrate that DeMix breaks the trade-off between sufficiency, accuracy and efficiency, obtaining the optimal mixture with higher benchmark performance at lower search cost. Additionally, we release the DeMix Corpora, a comprehensive 22T-token dataset comprising high-quality pre-training data with validated mixtures to facilitate open research. Our code and DeMix Corpora is available at https://github.com/Lucius-lsr/DeMix.

cs.CL↗

arXiv2026

Robust Tool Use via Fission-GRPO: Learning to Recover from Execution Errors

Large language models (LLMs) can call tools effectively, yet they remain brittle in multi-turn execution: after a tool-call error, smaller models often fall into repetitive invalid re-invocations instead of interpreting the feedback and recovering. This failure mode persists because current training paradigms do not explicitly teach models how to recover from execution errors. In particular, standard reinforcement learning (RL) collapses rich failure experience into sparse negative rewards, while pre-collected error-correction datasets become mismatched to the policy's evolving failure modes. To bridge this gap, we propose Fission-GRPO, a framework that converts execution errors into on-policy corrective supervision within the RL training loop. Our core mechanism fissions each failed trajectory into a new training instance by augmenting it with diagnostic feedback from a fine-tuned Error Simulator, then resampling multiple recovery rollouts on-policy. This enables the model to learn from the precise errors it makes during exploration, rather than from static, pre-collected error cases. On BFCL v4 Multi-Turn, Fission-GRPO improves the error recovery rate of Qwen3-8B by 5.7% absolute and overall accuracy by 4.0% (from 42.75% to 46.75%), outperforming both RL baselines and specialized tool-use agents. The method further generalizes to TAU-Bench and TAU2-Bench, achieving leading results across most settings with gains up to +17.4%.

cs.LG↗

arXiv2026

Unraveling Year-Long Radial Velocity Variations in Red Clump Region -- I: Comprehensive analysis of a K0 Giant star, 2 Draconis

Slow-rotating evolved stars frequently exhibit radial velocity (RV) variations on annual timescales, complicated by instrumental systematics and aliasing in the one-year regime. Here we investigate the origin of the near-yearly periodicity in 2 Dra, a star located in the red-clump region, assessing possible causes between stellar activity, instrumental profile (IP) effects, sampling alias, and planetary companions. We applied two independent approaches: (1) constraining diagnostic signals and performing a correlation analysis ($r$) between period-confined signals, and (2) evaluating phase stability by partitioning Keplerian fits. These methods enabled us to examine the physical connections and phase coherence among stellar activity indicators, RV measurements, and IP diagnostics. Our analysis suggests a stellar rotation period of $\simeq270\text{--}320$\,d for 2~Dra. The 340-d RV signal does not appear to originate from stellar activity in this chromospherically quiet star ($|r| \lesssim 0.33$), nor from instrumental systematics near the annual period ($|r| \lesssim 0.1$). This conclusion is supported by contrasting phase behavior: the RV and stellar activity phases remain stable, whereas the IP phases do not. We therefore propose that the 340-d variation likely arises from either small-amplitude intrinsic variability or a tentative gas giant companion with potential weak activity-induced modulation. The case of 2~Dra provides a framework for distinguishing the origins of $\sim$1-yr RV variations in other evolved stars.

astro-ph.SR↗

arXiv2025

A Tri-Dynamic Preprocessing Framework for UGC Video Compression

In recent years, user generated content (UGC) has become the dominant force in internet traffic. However, UGC videos exhibit a higher degree of variability and diverse characteristics compared to traditional encoding test videos. This variance challenges the effectiveness of data-driven machine learning algorithms for optimizing encoding in the broader context of UGC scenarios. To address this issue, we propose a Tri-Dynamic Preprocessing framework for UGC. Firstly, we employ an adaptive factor to regulate preprocessing intensity. Secondly, an adaptive quantization level is employed to fine-tune the codec simulator. Thirdly, we utilize an adaptive lambda tradeoff to adjust the rate-distortion loss function. Experimental results on large-scale test sets demonstrate that our method attains exceptional performance.

cs.MM↗

arXiv2025

A Preprocessing Framework for Video Machine Vision under Compression

There has been a growing trend in compressing and transmitting videos from terminals for machine vision tasks. Nevertheless, most video coding optimization method focus on minimizing distortion according to human perceptual metrics, overlooking the heightened demands posed by machine vision systems. In this paper, we propose a video preprocessing framework tailored for machine vision tasks to address this challenge. The proposed method incorporates a neural preprocessor which retaining crucial information for subsequent tasks, resulting in the boosting of rate-accuracy performance. We further introduce a differentiable virtual codec to provide constraints on rate and distortion during the training stage. We directly apply widely used standard codecs for testing. Therefore, our solution can be easily applied to real-world scenarios. We conducted extensive experiments evaluating our compression method on two typical downstream tasks with various backbone networks. The experimental results indicate that our approach can save over 15% of bitrate compared to using only the standard codec anchor version.

cs.MM↗

arXiv2025

RedOne 2.0: Rethinking Domain-specific LLM Post-Training in Social Networking Services

As a key medium for human interaction and information exchange, social networking services (SNS) pose unique challenges for large language models (LLMs): heterogeneous workloads, fast-shifting norms and slang, and multilingual, culturally diverse corpora that induce sharp distribution shift. Supervised fine-tuning (SFT) can specialize models but often triggers a ``seesaw'' between in-distribution gains and out-of-distribution robustness, especially for smaller models. To address these challenges, we introduce RedOne 2.0, an SNS-oriented LLM trained with a progressive, RL-prioritized post-training paradigm designed for rapid and stable adaptation. The pipeline consist in three stages: (1) Exploratory Learning on curated SNS corpora to establish initial alignment and identify systematic weaknesses; (2) Targeted Fine-Tuning that selectively applies SFT to the diagnosed gaps while mixing a small fraction of general data to mitigate forgetting; and (3) Refinement Learning that re-applies RL with SNS-centric signals to consolidate improvements and harmonize trade-offs across tasks. Across various tasks spanning three categories, our 4B scale model delivers an average improvements about 2.41 over the 7B sub-optimal baseline. Additionally, RedOne 2.0 achieves average performance lift about 8.74 from the base model with less than half the data required by SFT-centric method RedOne, evidencing superior data efficiency and stability at compact scales. Overall, RedOne 2.0 establishes a competitive, cost-effective baseline for domain-specific LLMs in SNS scenario, advancing capability without sacrificing robustness.

cs.AI↗

arXiv2025

Why Not Put a Microphone Near the Loudspeaker? A New Paradigm for Acoustic Echo Cancellation

Acoustic echo cancellation (AEC) remains challenging in real-world environments due to nonlinear distortions caused by low-cost loudspeakers and complex room acoustics. To mitigate these issues, we introduce a dual-microphone configuration, where an auxiliary reference microphone is placed near the loudspeaker to capture the nonlinearly distorted far-end signal. Although this reference signal is contaminated by near-end speech, we propose a preprocessing module based on Wiener filtering to estimate a compressed time-frequency mask to suppress near-end components. This purified reference signal enables a more effective linear AEC stage, whose residual error signal is then fed to a deep neural network for joint residual echo and noise suppression. Evaluation results show that our method outperforms baseline approaches on matched test sets. To evaluate its robustness under strong nonlinearities, we further test it on a mismatched dataset and observe that it achieves substantial performance gains. These results demonstrate its effectiveness in practical scenarios where the nonlinear distortions are typically unknown.

cs.SD↗

arXiv2025

RealBench: A Chinese Multi-image Understanding Benchmark Close to Real-world Scenarios

While various multimodal multi-image evaluation datasets have been emerged, but these datasets are primarily based on English, and there has yet to be a Chinese multi-image dataset. To fill this gap, we introduce RealBench, the first Chinese multimodal multi-image dataset, which contains 9393 samples and 69910 images. RealBench distinguishes itself by incorporating real user-generated content, ensuring high relevance to real-world applications. Additionally, the dataset covers a wide variety of scenes, image resolutions, and image structures, further increasing the difficulty of multi-image understanding. Ultimately, we conduct a comprehensive evaluation of RealBench using 21 multimodal LLMs of different sizes, including closed-source models that support multi-image inputs as well as open-source visual and video models. The experimental results indicate that even the most powerful closed-source models still face challenges when handling multi-image Chinese scenarios. Moreover, there remains a noticeable performance gap of around 71.8\% on average between open-source visual/video models and closed-source models. These results show that RealBench provides an important research foundation for further exploring multi-image understanding capabilities in the Chinese context.

cs.CL↗

arXiv2025

RedOne: Revealing Domain-specific LLM Post-Training in Social Networking Services

As a primary medium for modern information dissemination, social networking services (SNS) have experienced rapid growth, which has proposed significant challenges for platform content management and interaction quality improvement. Recently, the development of large language models (LLMs) has offered potential solutions but existing studies focus on isolated tasks, which not only encounter diminishing benefit from the data scaling within individual scenarios but also fail to flexibly adapt to diverse real-world context. To address these challenges, we introduce RedOne, a domain-specific LLM designed to break the performance bottleneck of single-task baselines and establish a comprehensive foundation for the SNS. RedOne was developed through a three-stage training strategy consisting of continue pretraining, supervised fine-tuning, and preference optimization, using a large-scale real-world dataset. Through extensive experiments, RedOne maintains strong general capabilities, and achieves an average improvement up to 14.02% across 8 major SNS tasks and 7.56% in SNS bilingual evaluation benchmark, compared with base models. Furthermore, through online testing, RedOne reduced the exposure rate in harmful content detection by 11.23% and improved the click page rate in post-view search by 14.95% compared with single-tasks finetuned baseline models. These results establish RedOne as a robust domain-specific LLM for SNS, demonstrating excellent generalization across various tasks and promising applicability in real-world scenarios.

cs.LG↗