arXiv subjects

Jian Yao

Publications and source records attributed to Jian Yao.

At least 19 recordsLinked to original sources

arXiv2026

Controlling Action-Head Co-Adaptation to VLM Representation Anisotropy

Modern learning systems increasingly combine pretrained models with downstream heads. Despite the growing importance of such modular designs, how the downstream model co-adapts to the geometry of pretrained representations remains relatively underexplored. We study this problem through vision-language-action models (VLAs), where a pretrained vision-language model (VLM) conditions an action model for control. We find that the action model progressively concentrates its sensitivity on high-variance dimensions of the VLM outputs. Through controlled experiments across different backbone-head models, we further show that counteracting this sensitivity concentration can be beneficial under representation--task mismatch settings. Motivated by these observations, we introduce \emph{Adaptive Interface Equalizer} (AdaIE), a parameter-free mechanism that reweights pretrained representations according to the evolving downstream sensitivity during training. AdaIE improves VLA success rates on simulation benchmarks, with gains of up to $+5.28$ percentage points (pp) on RoboCasa-GR1 and $+1.68$ pp on RoboTwin 2.0, and achieves an average gain of $+19$ pp across five real-world manipulation tasks. It further generalizes to diffusion policies and world models, improving downstream performance in both settings. Our results highlight interface co-adaptation as an important optimization problem when coupling pretrained representations with downstream models.

cs.AI↗

arXiv2026

Be Faithful When Response: Returning Fluent and Grounded Answers for Vision-Language Models Reinforcement Learning

Reinforcement Learning (RL) is an important paradigm for improving the reasoning capabilities of Vision-Language Models (VLMs). However, directly applying RL to rollout multimodal reasoning can lead to instability, due to the exploitation of language priors, the neglect of visual evidence, and the generation of reasoning traces that are fluent yet not visually grounded. The question arises: Can initially steer the policy toward visually faithful reasoning regime before applying reinforcement learning? To this end, we propose a Faithful Warm-Start (FWS) strategy that first curates samples with explicit vision-language causal relationships from six general VQA benchmarks to construct the FaithfulQA dataset, where each of the image-question pairs gains a certain degree of visual observations, question requirements, commonsense knowledge, domain knowledge, and the final answer. Subsequently, a VLM-based judge is employed to further purify the dataset, ensuring strong causal consistency and visual faithfulness. This warm-start stage equips the model with the capability to understand causally grounded vision-language patterns before subsequent RL optimization under sparse answer-level rewards. Experimental results show that such faithful supervision improves answer accuracy, stabilizes RL training, and reduces visually unsupported reasoning.

cs.AI↗

arXiv2026

Algorithm Modelization: Transforming Optimization Algorithms into Evolvable Intelligence

Can algorithms be learned and evolved as models, rather than repeatedly redesigned as programs? We explore this question by introducing Algorithm Modelization, a new paradigm that turns iterative algorithms into trainable models. Its central idea is to regard algorithm execution trajectories as computational experience: by learning how algorithmic states evolve over iterations, a model can internalize the knowledge embodied in existing algorithms, reproduce their computational behaviors, and further evolve beyond them through training. As a first realization of this paradigm, we modelize iterative optimization as an autoregressive process, where preceding generations constitute the computational context for generating the next population. We develop a dimension-independent population-temporal architecture to capture both population interactions and iterative dynamics, together with an imitation-to-surpassing learning framework that first acquires algorithmic knowledge from heterogeneous teachers and then improves the resulting optimizer directly through optimization feedback. The modelized optimizer consequently goes beyond reproducing individual algorithms: it consolidates heterogeneous search knowledge into a single trainable model, achieves stronger optimization capability through subsequent improvement, and generalizes its acquired knowledge across problems and substantially larger unseen dimensions. More fundamentally, Algorithm Modelization makes algorithmic capability accumulable through learning, where new algorithms and problem-solving experience can become training resources for progressively improving the model. Our results establish a promising path from manually designed algorithms toward general algorithm models that learn, generalize, and evolve from computational experience.

cs.AI↗

arXiv2026

CasePlay: Self-Play Reinforcement Learning from Case Reports for Medical Reasoning

Reinforcement learning (RL) is a promising way to improve the clinical reasoning abilities of large language models (LLMs), but current pipelines often rely on expert-labeled questions or costly synthetic-data generation. Meanwhile, the medical literature offers a wealth of high-quality unlabeled documents, including clinical case reports that trace the full clinical course from initial presentation to final treatment decisions. We study whether document-grounded self-play can convert these reports into effective RL training signals. We identify two challenges that limit existing self-play methods in this setting: knowledge-point fixation}, where the proposer repeatedly asks about a narrow set of salient clues in a source report, and difficulty-quality mismatch, where appropriately difficult questions may still be poorly formed or shallow, limiting their value for reasoner training. To address these challenges, we propose CasePlay, a self-play framework with two key components: (a) a Knowledge-Conditioned Proposer, which anchors question generation to case-specific knowledge points so that questions draw on a broader range of evidence from the source report; and (b) a Rubric-Judge Reasoner, which augments answer-accuracy feedback with rubric-based quality feedback on clue integration, reasoning depth, and option quality, guiding the proposer toward reasoning-intensive questions that better support reasoner training. Extensive experiments across seven medically related benchmarks demonstrate that CasePlay outperforms existing self-play baselines and improves the SFT warm-start model by 3.1 points on average, highlighted by a 6.3-point gain on LiveClin. This approach offers an effective path toward continually improving medical reasoning ability through self-play RL.

cs.LG↗

arXiv2026

Hierarchy of discriminative power and complexity in learning quantum ensembles

Distance metrics are central to machine learning, yet distances between ensembles of quantum states remain poorly understood due to fundamental quantum measurement constraints. We introduce a hierarchy of integral probability metrics, termed MMD-$k$, which generalizes the maximum mean discrepancy to quantum ensembles and exhibits a strict trade-off between discriminative power and statistical efficiency as the moment order $k$ increases. For pure-state ensembles of size $N$, estimating MMD-$k$ with arbitrary measurement schemes requires $Θ(N^{1-1/k})$ samples for constant $k$. At the same time, we prove that any stable distance metric with full discriminative power admits an $O(N\log N)$ upper bound and an $Ω(N)$ instance-gap lower bound. For quantum Wasserstein distance, with sufficiently large fixed state dimension, we establish a nearly linear lower bound in the ensemble size at constant additive accuracy, together with an $O(N\log N)$ upper bound. These results provide principled guidance for the design of loss functions in quantum machine learning, as we illustrate in training quantum denoising diffusion probabilistic models.

quant-ph↗

arXiv2026

IronLLM: Forging Compact Edge-Native Language Models for Real-Time Embodied Intelligence

We present IronLLM-0.6B, a 654M-parameter language model designed for efficient on-device inference. IronLLM-0.6B combines a hybrid attention architecture with X-MTP, a lightweight shared-KV multi-token prediction design that eliminates per-depth KV-cache replay and employs a lightweight verification head for rollback-free drafting, achieving a 1.48x decoding speedup. The model is pretrained on approximately 6.2 trillion tokens using a quality-oriented data pipeline and is further post-trained with Multi-Domain On-Policy Distillation to integrate capabilities from domain-specialized teachers. To better meet the low-latency requirements of on-device scenarios, IronLLM-0.6B adopts an Instruct-Only design. Evaluations show that IronLLM-0.6B achieves competitive performance relative to larger models such as Qwen3.5-0.8B and MiniCPM5-1B, while producing more concise responses on many tasks. We further present IronLLM-0.6B-Light, which replaces RMSNorm with Dynamic Tanh and simplifies several computationally expensive components to improve inference and quantization efficiency. Together, the IronLLM models provide an effective performance-efficiency trade-off for resource-constrained deployment.

cs.AI↗

arXiv2026

IronViT: Toward Efficient Generalist Visual Representation Learning

A generalist vision encoder must capture semantic, spatial, language-aligned, and action-relevant cues within a unified representation, yet softmax attention underlying today's most capable visual backbones becomes prohibitively expensive at high resolution. A natural attempt to address both challenges is to distill multiple specialist teachers directly into an efficient architecture. We find that directly coupling these objectives degrades representation quality, as the student must simultaneously reconcile heterogeneous capabilities and adapt them to a different token-mixing architecture. We introduce IronViT, built on a simple principle: consolidate capabilities before constraining computation. IronViT first distills complementary specialists into a softmax attention capability bridge, then progressively transfers the consolidated representation to a hybrid softmax-linear attention encoder. A purpose-built data pipeline further curates the distillation corpus for higher information density and broader domain coverage. Across recognition, retrieval, dense prediction, multimodal understanding, and robotic learning, IronViT is competitive with leading specialist and generalist vision encoders. The softmax bridge achieves the strongest aggregate performance in multimodal understanding and robotic learning among the evaluated backbones, while the hybrid encoder retains broad transfer performance with an efficiency advantage that grows with input resolution. Together, these results show that consolidating capabilities before architectural conversion can yield a generalist visual encoder without inheriting the prohibitive high-resolution cost of conventional softmax attention.

cs.CV↗

arXiv2026

RoofSeg: An edge-aware transformer-based network for end-to-end roof plane segmentation

Roof plane segmentation is one of the key procedures for reconstructing three-dimensional (3D) building models at levels of detail (LoD) 2 and 3 from airborne light detection and ranging (LiDAR) point clouds. The majority of current approaches for roof plane segmentation rely on the manually designed or learned features followed by some specifically designed geometric clustering strategies. Because the learned features are more powerful than the manually designed features, the deep learning-based approaches usually perform better than the traditional approaches. However, the current deep learning-based approaches have three unsolved problems. The first is that most of them are not truly end-to-end, the plane segmentation results may be not optimal. The second is that the point feature discriminability near the edges is relatively low, leading to inaccurate planar edges. The third is that the planar geometric characteristics are not sufficiently considered to constrain the network training. To solve these issues, a novel edge-aware transformer-based network, named RoofSeg, is developed for segmenting roof planes from LiDAR point clouds in a truly end-to-end manner. In the RoofSeg, we leverage a transformer encoder-decoder-based framework to hierarchically predict the plane instance masks with the use of a set of learnable plane queries. To further improve the segmentation accuracy of edge regions, we also design an Edge-Aware Mask Module (EAMM) that sufficiently incorporates planar geometric prior of edges to enhance its discriminability for plane instance mask refinement. In addition, we propose an adaptive weighting strategy in the mask loss to reduce the influence of misclassified points, and also propose a new plane geometric loss to constrain the network training.

cs.CV↗

arXiv2026

Beyond Retrieval: Learning Compact User Representations for Scalable LLM Personalization

Personalizing large language models requires adapting model behavior to individual users while preserving robustness and deployment-scale efficiency. Existing approaches typically personalize LLMs either at the input level, by retrieving user histories or constructing profile prompts, or at the parameter level, by maintaining user-specific parameter-efficient modules. The former makes personalization sensitive to retrieval quality and prompt design, whereas the latter incurs storage and maintenance costs that grow with the user population. To address these limitations, we propose TAP-PER (Temporal Attentive Prefix for PERsonalization), a prefix-based framework that encodes user preferences as learnable representations, avoiding the serialization of user histories into prompts and replacing heavy per-user adapters with lightweight user-state prefix embeddings. Inspired by personalized recommendation systems, TAP-PER decomposes user modeling into user-state and query-conditioned components, and incorporates temporal signals to capture the evolving nature of user interests. Experiments on six LaMP tasks show that TAP-PER consistently outperforms prompt-based and model-based baselines across classification, rating, and generation settings. Moreover, TAP-PER uses 130x fewer per-user parameters than OPPU and roughly half the total parameter footprint of PER-PCS at the 1,000-user scale, demonstrating scalable personalization without prompt-serialized histories or heavy per-user adapters.

cs.IR↗

arXiv2026

Hypothesis testing between quantum ensembles

Quantum state ensembles are important in quantum information processing. For example, quantum $t$-designs model highly entangled states in complex systems, while projected ensembles appear in generative quantum machine learning and studies of thermalization. With their sample state accompanied by a classical label, these ensembles contain operational information beyond their average density operators. Yet an ensemble differs from a classical-quantum state because it is invariant under permutations of labels. We formulate binary hypothesis testing between finite quantum ensembles and derive fundamental limits on error probability. Given an observed label pattern, we show that the joint sampled state can be described by power-weighted ensemble moments. This yields the Bayes-optimal measurement and exact finite-sample error, revealing that discrimination is governed by the full moment hierarchy up to the number of samples. In the many-sample limit, we derive Chernoff bounds and obtain exact error exponents for finite uniform pure-state ensembles. We apply these results to optical communication and $t$-designs. For finite uniform pure-state $t$-designs with large $t$, the maximal discrimination exponent scales sharply as $\sim t^{-2}$, while equal-prior fixed-error testing requires $\sim t^2$ samples.

quant-ph↗

arXiv2026

SLAT: Segment-Level Adaptive Trimming for Efficient CoT Reasoning

Recent advances in Large Reasoning Models have significantly improved chain-of-thought (CoT) capabilities via reinforcement learning (RL). However, generated reasoning chains frequently suffer from structural redundancy (i.e., \emph{overthinking}), incurring high computational overhead without improving answer correctness. Existing mitigation strategies typically rely on token-uniform length penalties, which provide coarse, segment-agnostic pressure toward shorter outputs and can inadvertently suppress useful reasoning alongside redundancy. To address this, we demonstrate that inefficiency concentrates in high-probability segments with low marginal utility. We derive a theoretical characterization of segment suboptimality under the correctness-length trade-off objective and propose \textsc{SLAT} (Segment-Level Adaptive Trimming), an RL framework that selectively suppresses redundant segments based on this criterion. Empirical results on standard benchmarks indicate that \textsc{SLAT} establishes a superior accuracy-efficiency Pareto frontier, reducing reasoning length by $50\%$ relative to uncompressed baselines while maintaining competitive accuracy. Overall, our results suggest that theoretically grounded, segment-aware trimming is a promising direction for efficient CoT reasoning in large language models.

cs.AI↗

arXiv2026

SpatialForge: Bootstrapping 3D-Aware Spatial Reasoning from Open-World 2D Images

Recent advancements in Large Vision-Language Models (VLMs) have demonstrated exceptional semantic understanding, yet these models consistently struggle with spatial reasoning, often failing at fundamental geometric tasks such as depth ordering and precise coordinate grounding. Recent efforts introduce spatial supervision from scene-centric datasets (e.g., multi-view scans or indoor video), but are constrained by the limited number of underlying scenes. As a result, the scale and diversity of such data remain significantly smaller than those of web-scale 2D image collections. To address this limitation, we propose SpatialForge, a scalable data synthesis pipeline that transforms in-the-wild 2D images into spatial reasoning supervision. Our approach decomposes spatial reasoning into perception and relation, and constructs structured supervision signals covering depth, layout, and viewpoint-dependent reasoning, with automatic verification to ensure data quality. Based on this pipeline, we build SpatialForge-10M, a large-scale dataset containing 10 million spatial QA pairs. Extensive experiments across multiple spatial reasoning benchmarks demonstrate that training on SpatialForge-10M significantly improves the spatial reasoning ability of standard VLMs, highlighting the effectiveness of scaling 2D data for 3D-aware spatial reasoning.

cs.CV↗

arXiv2026

Beyond Single-Sample: Reliable Multi-Sample Distillation for Video Understanding

Traditional black-box distillation for Large Vision-Language Models (LVLMs) typically relies on a single teacher response per input, which often yields high-variance responses and format inconsistencies in multimodal or temporal scenarios. To mitigate this unreliable supervision, we propose R-MSD (Reliable Multi-Sample Distillation), a framework that explicitly models teacher sampling variance to enhance distillation stability. Rather than relying on a single teacher response, our approach leverages a task-adaptive teacher pool to provide robust supervision tailored to both closed-ended and open-ended reasoning. By integrating quality-aware signal matching with an adversarial distillation objective, our approach effectively filters teacher noise while maximizing knowledge transfer. Extensive evaluations across comprehensive video understanding benchmarks demonstrate that R-MSD consistently outperforms single sample distillation methods. We additionally include an original SFT+RL 4B baseline under the same training budget, which shows only marginal gains, while our method achieves significant improvements. With a 4B student model, our approach delivers gains on VideoMME (+1.5%), Video-MMMU (+3.2%), and MathVerse (+3.6%).

cs.CV↗

arXiv2026

VAR-MATH: Probing True Mathematical Reasoning in LLMS via Symbolic Multi-Instance Benchmarks

Recent advances in reinforcement learning (RL) have led to substantial improvements in the mathematical reasoning abilities of LLMs, as measured by standard benchmarks. Yet these gains often persist even when models are trained with flawed signals, such as random or inverted rewards. This raises a fundamental question: do such improvements reflect genuine reasoning, or are they merely artifacts of overfitting to benchmark-specific patterns? To answer this question, we adopt an evaluation-centric perspective and highlight two critical shortcomings in existing protocols. First, benchmark contamination arises because test problems are publicly available, thereby increasing the risk of data leakage. Second, evaluation fragility results from reliance on single-instance assessments, which are sensitive to stochastic outputs and fail to capture reasoning consistency. These limitations suggest the need for a new evaluation paradigm that can probe reasoning ability beyond memorization and one-off success. As response, we propose VAR-MATH, a symbolic evaluation framework that converts fixed numerical problems into parameterized templates and requires models to solve multiple instantiations of each. This design enforces consistency across structurally equivalent variants, mitigates contamination, and enhances robustness through bootstrapped metrics. We apply VAR-MATH to transform three popular benchmarks, AMC23, AIME24, and AIME25, into their symbolic counterparts, VAR-AMC23, VAR-AIME24, and VAR-AIME25. Experimental results show substantial performance drops for RL-trained models on these variabilized benchmarks, especially for smaller models, with average declines of 47.9\% on AMC23, 58.8\% on AIME24, and 72.9\% on AIME25. These findings indicate that some existing RL methods rely on superficial heuristics and fail to generalize beyond specific numerical forms.

cs.AI↗

arXiv2025

Full-sky Models of Galactic Microwave Emission and Polarization at Sub-arcminute Scales for the Python Sky Model

Polarized foreground emission from the Galaxy is one of the biggest challenges facing current and upcoming cosmic microwave background (CMB) polarization experiments. We develop new models of polarized Galactic dust and synchrotron emission at CMB frequencies that draw on the latest observational constraints, that employ the ``polarization fraction tensor'' framework to couple intensity and polarization in a physically motivated way, and that allow for stochastic realizations of small-scale structure at sub-arcminute angular scales currently unconstrained by full-sky data. We implement these models into the publicly available Python Sky Model (PySM) software and additionally provide PySM interfaces to select models of dust and CO emission from the literature. We characterize the behavior of each model by quantitatively comparing it to observational constraints in both maps and power spectra, demonstrating an overall improvement over previous PySM models. Finally, we synthesize models of the various Galactic foreground components into a coherent suite of three plausible microwave skies that span a range of astrophysical complexity allowed by current data.

astro-ph.CO↗

arXiv2025

Diversity-Aware Policy Optimization for Large Language Model Reasoning

The reasoning capabilities of large language models (LLMs) have advanced rapidly, particularly following the release of DeepSeek R1, which has inspired a surge of research into data quality and reinforcement learning (RL) algorithms. Despite the pivotal role diversity plays in RL, its influence on LLM reasoning remains largely underexplored. To bridge this gap, this work presents a systematic investigation into the impact of diversity in RL-based training for LLM reasoning, and proposes a novel diversity-aware policy optimization method. Across evaluations on 12 LLMs, we observe a strong positive correlation between the solution diversity and Potential at k (a novel metric quantifying an LLM's reasoning potential) in high-performing models. This finding motivates our method to explicitly promote diversity during RL training. Specifically, we design a token-level diversity and reformulate it into a practical objective, then we selectively apply it to positive samples. Integrated into the R1-zero training framework, our method achieves a 3.5 percent average improvement across four mathematical reasoning benchmarks, while generating more diverse and robust solutions.

cs.LG↗

arXiv2025

HRT1: One-Shot Human-to-Robot Trajectory Transfer for Mobile Manipulation

We introduce a novel system for human-to-robot trajectory transfer that enables robots to manipulate objects by learning from human demonstration videos. The system consists of four modules. The first module is a data collection module that is designed to collect human demonstration videos from the point of view of a robot using an AR headset. The second module is a video understanding module that detects objects and extracts 3D human-hand trajectories from demonstration videos. The third module transfers a human-hand trajectory into a reference trajectory of a robot end-effector in 3D space. The last module utilizes a trajectory optimization algorithm to solve a trajectory in the robot configuration space that can follow the end-effector trajectory transferred from the human demonstration. Consequently, these modules enable a robot to watch a human demonstration video once and then repeat the same mobile manipulation task in different environments, even when objects are placed differently from the demonstrations. Experiments of different manipulation tasks are conducted on a mobile manipulator to verify the effectiveness of our system

cs.RO↗

arXiv2025

DFG-PCN: Point Cloud Completion with Degree-Flexible Point Graph

Point cloud completion is a vital task focused on reconstructing complete point clouds and addressing the incompleteness caused by occlusion and limited sensor resolution. Traditional methods relying on fixed local region partitioning, such as k-nearest neighbors, which fail to account for the highly uneven distribution of geometric complexity across different regions of a shape. This limitation leads to inefficient representation and suboptimal reconstruction, especially in areas with fine-grained details or structural discontinuities. This paper proposes a point cloud completion framework called Degree-Flexible Point Graph Completion Network (DFG-PCN). It adaptively assigns node degrees using a detail-aware metric that combines feature variation and curvature, focusing on structurally important regions. We further introduce a geometry-aware graph integration module that uses Manhattan distance for edge aggregation and detail-guided fusion of local and global features to enhance representation. Extensive experiments on multiple benchmark datasets demonstrate that our method consistently outperforms state-of-the-art approaches.

cs.GR↗