arXiv subjects

Jinghan Zhang

Publications and source records attributed to Jinghan Zhang.

At least 19 recordsLinked to original sources

arXiv2026

BrailleBench: Investigating Multi-Criteria Braille Comprehension in Large Language Models

Although Large language models (LLMs) mediate access to knowledge and computational assistance, their capabilities should benefit vulnerable groups in the same way. However, it is unclear whether existing AI systems are inclusive enough for blind and deafblind users to access the same functionality through Braille, whose indicators, contractions, and digital representations introduce distinct requirements for model comprehension. To this end, we introduce BrailleBench, a benchmark for evaluating LLMs in Braille comprehension from different Criteria. BrailleBench aligns 5,570 instances from five datasets, including mathematics, commonsense, and multi-hop question answering across English and Braille Grades 1 and 2. Different configurations are designed to understand whether the systems can comprehend Braille-authored content, express answers in Braille, and complete end-to-end Braille interaction. To ensure the quality and prevent evaluation bias, the benchmark is built through a deterministic, expert-reviewed pipeline via a self-created Braille Toolkit without using any data instances generated by LLMs. We evaluate six representative LLMs from various aspects. The results reveal a persistent gap between print-English capability and Braille accessibility. Braille understanding and expression are asymmetric, where Grade 2 is especially fragile on the input side compared to Grade 1, and fully Braille requests further reduce performance. The experimental observations provide valuable guidance for the development of future Braille AI systems. All related resources in BrailleBench are publicly available for future research.

cs.AI↗

arXiv2026

The Generalization Spectrum: A Chromatographic Approach to Evaluating Learning Algorithms

Traditional evaluations measure a learning algorithm's final performance on an i.i.d. test set, reducing learning to a single aggregate score. This approach obscures a fundamental question: to what extent does learning from a specific example generalize to others? Such per-sample generalization, akin to learning by analogy in human cognition, captures how far the knowledge extracted from one example can transfer, yet remains invisible to standard benchmarks. We introduce the Generalization Spectrum, an evaluation framework designed to expose this hidden dimension. For each training example, we construct a controlled suite of test variants arranged by increasing transfer distance, from exact recall to implementation transfer across languages, context transfer under complete narrative re-framing, category-matched in-domain problems, and an unpaired baseline. By tracking performance across these distances, we reveal not just whether an algorithm learns, but how far that learning extends. We instantiate this framework on competitive programming, using a selection-and-synthesis pipeline seeded with recent problems to mitigate contamination. We first compare three canonical learning paradigms under matched memorization. RL converts memorization into near-transfer more efficiently than SFT-family baselines, while ICL exhibits strong but correspondence-dependent transfer. We then use the Spectrum to diagnose within-family variants. The resulting profiles show that local gains need not expand the generalization radius: abstractions and hints mainly lift local transfer, RFT preserves a stronger far-transfer tail than reference SFT, and self-distillation or hint-assisted RL can reduce far transfer even when local transfer or optimization improves.

cs.LG↗

arXiv2026

OctCGS: Octree-Contextual Gaussian Splatting with Explicit Multi-Order Propagation Modeling for Channel Knowledge Map Construction

Channel knowledge maps (CKMs) learn the relation between transmitter (Tx) and receiver (Rx) positions and channel knowledge to support environment-aware wireless communications. Implicit neural methods can model continuous channel variation but often incur high training and inference cost, while existing Gaussian-splatting-based CKM methods improve efficiency yet still compress wireless multipath interactions into aggregated scattering representations. Consequently, explicit modeling of multi-bounce wireless propagation remains absent from CKM construction. We propose OctCGS, an octree-contextual Gaussian splatting framework that explicitly models the order of bounce jointly over Tx/Rx positions and carrier frequencies. OctCGS partitions the environment into a multi-resolution octree and anchors one Gaussian primitive to each leaf node. Rather than having each Gaussian independently encode all multi-path propagations, it models complex electromagnetic interactions among scatterers through tree attention over the octree hierarchy with controlled complexity. Experiments on simulated benchmarks show that OctCGS achieves a 2.99 dB channel-gain mean absolute error (MAE) and 0.065 channel gain normalized mean absolute error (NMAE), outperforming the strongest baseline by 0.88 dB MAE and 0.021 NMAE.

eess.SP↗

arXiv2026

Planar Gaussian Splatting with Bilinear Spatial Transformer for Wireless Radiance Field Reconstruction

Wireless radiance field (WRF) reconstruction aims to learn a continuous, queryable representation of radio frequency characteristics over 3D space and direction, from which specific quantities, such as the spatial power spectrum (SPS) at a receiver given a transmitter position, can be predicted. While Gaussian splatting (GS)-based method has surpassed Neural Radiance Fields (NeRF)-based method for this task, existing adaptations largely transplant vision pipelines, limiting physical interpretability and accuracy. We introduce BiSplat-WRF, a planar GS framework that retains the expressiveness of 3D GS while removing unnecessary projections and incorporating global EM coupling and mutual scattering among primitives. Each primitive is a 2D planar Gaussian with 3D coordinates, rendered directly on the angular domain of the SPS. A bilinear spatial transformer (BST) aggregates inter-primitive relations on an angular grid and, via attention, captures long-range electromagnetic dependencies, thereby enforcing globally aware EM interactions that reflect the complex physics of the wireless environment. On spatial spectrum synthesis task, BiSplat-WRF surpasses NeRF-based and prior GS-based baselines with respect to the Structural Similarity Index (SSIM); comprehensive ablation studies validate the contribution of BST. We also provide a larger BiSplat-WRF+ variant that further increases SSIM at a higher computation cost, serving as a strong reference for future studies.

eess.SP↗

arXiv2026

StaRPO: Stability-Augmented Reinforcement Policy Optimization

Reinforcement learning (RL) is effective in enhancing the accuracy of large language models in complex reasoning tasks. Existing RL policy optimization frameworks rely on final-answer correctness as feedback signals and rarely capture the internal logical structure of the reasoning process. Consequently, the models would generate fluent and semantically relevant responses but logically inconsistent, structurally erratic, or redundant. To this end, we propose StaRPO, a stability-augmented reinforcement learning framework that explicitly incorporates reasoning stability into the optimization objective. Our StaRPO decomposes stability into two computable lightweight metrics: the Autocorrelation Function (ACF) to evaluate local step-to-step coherence, and Path Efficiency (PE) to evaluate global goal-directedness of the reasoning trajectory. These stability rewards are combined with task rewards to provide complementary and process-aware feedback. We validate the effectiveness of using ACF and PE rewards by showing their correlation with logic errors on two backbone models. Experiments on four reasoning benchmarks show that StaRPO consistently outperforms compared baselines and can enhance both final-answer accuracy and logical stability.

cs.AI↗

arXiv2026

Sim2Act: Robust Simulation-to-Decision Learning via Adversarial Calibration and Group-Relative Perturbation

Simulation-to-decision learning enables safe policy training in digital environments without risking real-world deployment, and has become essential in mission-critical domains such as supply chains and industrial systems. However, simulators learned from noisy or biased real-world data often exhibit prediction errors in decision-critical regions, leading to unstable action ranking and unreliable policies. Existing approaches either focus on improving average simulation fidelity or adopt conservative regularization, which may cause policy collapse by discarding high-risk high-reward actions. We propose Sim2Act, a robust simulation-to-decision framework that addresses both simulator and policy robustness. First, we introduce an adversarial calibration mechanism that re-weights simulation errors in decision-critical state-action pairs to align surrogate fidelity with downstream decision impact. Second, we develop a group-relative perturbation strategy that stabilizes policy learning under simulator uncertainty without enforcing overly pessimistic constraints. Extensive experiments on multiple supply chain benchmarks demonstrate improved simulation robustness and more stable decision performance under structured and unstructured perturbations.

cs.LG↗

arXiv2026

SkillCraft: Can LLM Agents Learn to Use Tools Skillfully?

Real-world tool-using agents operate over long-horizon workflows with recurring structure and diverse demands, where effective behavior requires not only invoking atomic tools but also abstracting, and reusing higher-level tool compositions. However, existing benchmarks mainly measure instance-level success under static tool sets, offering limited insight into agents' ability to acquire such reusable skills. We address this gap by introducing SkillCraft, a benchmark explicitly stress-test agent ability to form and reuse higher-level tool compositions, where we call Skills. SkillCraft features realistic, highly compositional tool-use scenarios with difficulty scaled along both quantitative and structural dimensions, designed to elicit skill abstraction and cross-task reuse. We further propose a lightweight evaluation protocol that enables agents to auto-compose atomic tools into executable Skills, cache and reuse them inside and across tasks, thereby improving efficiency while accumulating a persistent library of reusable skills. Evaluating state-of-the-art agents on SkillCraft, we observe substantial efficiency gains, with token usage reduced by up to 80% by skill saving and reuse. Moreover, success rate strongly correlates with tool composition ability at test time, underscoring compositional skill acquisition as a core capability.

cs.CL↗

arXiv2026

Towards Dynamic Dense Retrieval with Routing Strategy

The \textit{de facto} paradigm for applying dense retrieval (DR) to new tasks involves fine-tuning a pre-trained model for a specific task. However, this paradigm has two significant limitations: (1) It is difficult adapt the DR to a new domain if the training dataset is limited. (2) Old DR models are simply replaced by newer models that are trained from scratch when the former are no longer up to date. Especially for scenarios where the model needs to be updated frequently, this paradigm is prohibitively expensive. To address these challenges, we propose a novel dense retrieval approach, termed \textit{dynamic dense retrieval} (DDR). DDR uses \textit{prefix tuning} as a \textit{module} specialized for a specific domain. These modules can then be compositional combined with a dynamic routing strategy, enabling highly flexible domain adaptation in the retrieval part. Extensive evaluation on six zero-shot downstream tasks demonstrates that this approach can surpass DR while utilizing only 2\% of the training parameters, paving the way to achieve more flexible dense retrieval in IR. We see it as a promising future direction for applying dense retrieval to various tasks.

cs.IR↗

arXiv2026

OpenDecoder: Open Large Language Model Decoding to Incorporate Document Quality in RAG

The development of large language models (LLMs) has achieved superior performance in a range of downstream tasks, including LLM-based retrieval-augmented generation (RAG). The quality of generated content heavily relies on the usefulness of the retrieved information and the capacity of LLMs' internal information processing mechanism to incorporate it in answer generation. It is generally assumed that the retrieved information is relevant to the question. However, the retrieved information may have a variable degree of relevance and usefulness, depending on the question and the document collection. It is important to take into account the relevance of the retrieved information in answer generation. In this paper, we propose OpenDecoder, a new approach that leverages explicit evaluation of the retrieved information as quality indicator features for generation. We aim to build a RAG model that is more robust to varying levels of noisy context. Three types of explicit evaluation information are considered: relevance score, ranking score, and QPP (query performance prediction) score. The experimental results on five benchmark datasets demonstrate the effectiveness and better robustness of OpenDecoder by outperforming various baseline methods. Importantly, this paradigm is flexible to be integrated with the post-training of LLMs for any purposes and incorporated with any type of external indicators.

cs.CL↗

arXiv2025

Parametric Retrieval-Augmented Generation using Latent Routing of LoRA Adapters

Parametric Retrieval-Augmented Generation (PRAG) is a RAG approach that integrates external knowledge directly into model parameters using a LoRA adapter, aiming at reducing the inference cost compared to traditional RAG. However, current PRAG approaches adopt a \textit{one-to-one} document encoding scheme, using a dedicated LoRA adapter for each individual document. This scheme introduces two major limitations: 1) As the number of documents increases, there will be a prohibitive cost for training and storage. 2) The LoRA adapters may largely overlap due to the shared knowledge across documents, making the approach highly inefficient. To overcome these challenges, we propose the Poly-PRAG approach, which uses a small set of LoRA adapters that are able to encode more general knowledge. Each document can be encoded using a combination of them through a latent routing function. By jointly training the LoRA adapters and the latent routing function, each LoRA adapter is able to encode a shared part of the knowledge across documents, and the routing function can select the best combination of adapters for a document. Experimental results on four benchmarks demonstrate the effectiveness of the Poly-PRAG compared to other strong PRAG baselines. In addition, this approach reduces the storage requirement by avoiding the need to store a large number of LoRA adapters and offers a more efficient way to encode external knowledge into LLMs.

cs.IR↗

arXiv2025

Retrieval-Augmented Feature Generation for Domain-Specific Classification

Feature generation can significantly enhance learning outcomes, particularly for tasks with limited data. An effective way to improve feature generation is to expand the current feature space using existing features and enriching the informational content. However, generating new, interpretable features usually requires domain-specific knowledge on top of the existing features. In this paper, we introduce a Retrieval-Augmented Feature Generation method, RAFG, to generate useful and explainable features specific to domain classification tasks. To increase the interpretability of the generated features, we conduct knowledge retrieval among the existing features in the domain to identify potential feature associations. These associations are expected to help generate useful features. Moreover, we develop a framework based on large language models (LLMs) for feature generation with reasoning to verify the quality of the features during their generation process. Experiments across several datasets in medical, economic, and geographic domains show that our RAFG method can produce high-quality, meaningful features and significantly improve classification performance compared with baseline methods.

cs.CL↗

arXiv2025

Blind Spot Navigation in Large Language Model Reasoning with Thought Space Explorer

Large language models have shown strong reasoning capabilities through chain-structured methods such as Chain-of-Thought. Recent studies optimize thought structures by generating parallel or tree-like structures, switching between long and short reasoning modes, or aligning reasoning steps with task performance. However, these approaches mainly rely on previously generated logical directions of the chains, which ignore the unexplored regions of the solution space. Such a phenomenon is defined as blind spots, which limit the diversity and effectiveness of the reasoning process. To this end, we propose the ``Thought Space Explorer'' (TSE), a framework for navigating and expanding thought structures to overcome blind spots in LLM reasoning. Our TSE first identifies key nodes with high impact, then generates new nodes by integrating information from multiple chains. Finally, it extends new branches through connection strategies. We conduct a series of experiments on math and QA benchmarks. Compared with existing baseline methods, TSE improves the accuracy of both the final answer and intermediate reasoning steps, while maintaining a better effectiveness-efficiency trade-off for practical deployment.

cs.CL↗

arXiv2025

Why Do LLM Agents Fail in Exploring New Environments? A World-Modeling Perspective

Large Language Models (LLMs) as agents often fail to improve in new environments. We identify and characterize a failure mode we call exploration collapse: under reinforcement learning (RL) in environments whose states are unfamiliar to the policy, Pass@k, the probability that at least one of k sampled trajectories succeeds, drops markedly over training even as Pass@1 edges up, revealing increasingly brittle exploration; environments closer to the pretraining distribution show no such decline. We trace this collapse to weak grounding in environment states and dynamics, and study a simple remedy: explicitly teaching the agent to estimate the current state and predict its transitions before optimizing for reward. We instantiate it as SPA, an explore-then-exploit recipe that cold-starts the policy with a Self-Experience supervised finetuning (SFT) stage, collecting the model's own interaction trajectories and supervising state and next-state prediction, and then runs standard RL. The resulting world model serves as a grounded initialization for RL rather than an inference-time planner. Across unseen environments, SPA consistently and substantially improves over vanilla RL: for example, it raises the Sokoban success rate from 25.6% to 59.8% on Qwen2.5-1.5B-Instruct, letting sub-3B models surpass a 20B baseline on these tasks. Controlled studies indicate that the gains track four factors: grounded state representations, explicit transition modeling, self-experience trajectories from a sufficiently strong exploration policy, and adequate coverage of transition data.

cs.LG↗

arXiv2025

Why Is Spatial Reasoning Hard for VLMs? An Attention Mechanism Perspective on Focus Areas

Large Vision Language Models (VLMs) have long struggled with spatial reasoning tasks. Surprisingly, even simple spatial reasoning tasks, such as recognizing "under" or "behind" relationships between only two objects, pose significant challenges for current VLMs. In this work, we study the spatial reasoning challenge from the lens of mechanistic interpretability, diving into the model's internal states to examine the interactions between image and text tokens. By tracing attention distribution over the image through out intermediate layers, we observe that successful spatial reasoning correlates strongly with the model's ability to align its attention distribution with actual object locations, particularly differing between familiar and unfamiliar spatial relationships. Motivated by these findings, we propose ADAPTVIS based on inference-time confidence scores to sharpen the attention on highly relevant regions when confident, while smoothing and broadening the attention window to consider a wider context when confidence is lower. This training-free decoding method shows significant improvement (e.g., up to a 50 absolute point improvement) on spatial reasoning benchmarks such as WhatsUp and VSR with negligible cost. We make code and data publicly available for research purposes at https://github.com/shiqichen17/AdaptVis.

cs.CL↗

arXiv2025

Data-Efficient Symbolic Regression via Foundation Model Distillation

Discovering interpretable mathematical equations from observed data (a.k.a. equation discovery or symbolic regression) is a cornerstone of scientific discovery, enabling transparent modeling of physical, biological, and economic systems. While foundation models pre-trained on large-scale equation datasets offer a promising starting point, they often suffer from negative transfer and poor generalization when applied to small, domain-specific datasets. In this paper, we introduce EQUATE (Equation Generation via QUality-Aligned Transfer Embeddings), a data-efficient fine-tuning framework that adapts foundation models for symbolic equation discovery in low-data regimes via distillation. EQUATE combines symbolic-numeric alignment with evaluator-guided embedding optimization, enabling a principled embedding-search-generation paradigm. Our approach reformulates discrete equation search as a continuous optimization task in a shared embedding space, guided by data-equation fitness and simplicity. Experiments across three standard public benchmarks (Feynman, Strogatz, and black-box datasets) demonstrate that EQUATE consistently outperforms state-of-the-art baselines in both accuracy and robustness, while preserving low complexity and fast inference. These results highlight EQUATE as a practical and generalizable solution for data-efficient symbolic regression in foundation model distillation settings.

cs.LG↗

arXiv2025

Galaxy: A Cognition-Centered Framework for Proactive, Privacy-Preserving, and Self-Evolving LLM Agents

Intelligent personal assistants (IPAs) such as Siri and Google Assistant are designed to enhance human capabilities and perform tasks on behalf of users. The emergence of LLM agents brings new opportunities for the development of IPAs. While responsive capabilities have been widely studied, proactive behaviors remain underexplored. Designing an IPA that is proactive, privacy-preserving, and capable of self-evolution remains a significant challenge. Designing such IPAs relies on the cognitive architecture of LLM agents. This work proposes Cognition Forest, a semantic structure designed to align cognitive modeling with system-level design. We unify cognitive architecture and system design into a self-reinforcing loop instead of treating them separately. Based on this principle, we present Galaxy, a framework that supports multidimensional interactions and personalized capability generation. Two cooperative agents are implemented based on Galaxy: KoRa, a cognition-enhanced generative agent that supports both responsive and proactive skills; and Kernel, a meta-cognition-based meta-agent that enables Galaxy's self-evolution and privacy preservation. Experimental results show that Galaxy outperforms multiple state-of-the-art benchmarks. Ablation studies and real-world interaction cases validate the effectiveness of Galaxy.

cs.AI↗

arXiv2025

Tensorized Clustered LoRA Merging for Multi-Task Interference

Despite the success of the monolithic dense paradigm of large language models (LLMs), the LoRA adapters offer an efficient solution by fine-tuning small task-specific modules and merging them with the base model. However, in multi-task settings, merging LoRA adapters trained on heterogeneous sources frequently causes \textit{task interference}, degrading downstream performance. To address this, we propose a tensorized clustered LoRA (TC-LoRA) library targeting to address the task interference at the \textit{text-level} and \textit{parameter-level}. At the \textit{text-level}, we cluster the training samples in the embedding space to capture input-format similarities, then train a specialized LoRA adapter for each cluster. At the \textit{parameter-level}, we introduce a joint Canonical Polyadic (CP) decomposition that disentangles task-specific and shared factors across LoRA adapters. This joint factorization preserves essential knowledge while reducing cross-task interference. Extensive experiments on out-of-domain zero-shot and skill-composition tasks-including reasoning, question answering, and coding. Compared to strong SVD-based baselines, TC-LoRA achieves +1.4\% accuracy on Phi-3 and +2.3\% on Mistral-7B (+2.3\%), demonstrating the effectiveness of TC-LoRA in LLM adaptation.

cs.LG↗

arXiv2025

ConvMix: A Mixed-Criteria Data Augmentation Framework for Conversational Dense Retrieval

Conversational search aims to satisfy users' complex information needs via multiple-turn interactions. The key challenge lies in revealing real users' search intent from the context-dependent queries. Previous studies achieve conversational search by fine-tuning a conversational dense retriever with relevance judgments between pairs of context-dependent queries and documents. However, this training paradigm encounters data scarcity issues. To this end, we propose ConvMix, a mixed-criteria framework to augment conversational dense retrieval, which covers more aspects than existing data augmentation frameworks. We design a two-sided relevance judgment augmentation schema in a scalable manner via the aid of large language models. Besides, we integrate the framework with quality control mechanisms to obtain semantically diverse samples and near-distribution supervisions to combine various annotated data. Experimental results on five widely used benchmarks show that the conversational dense retriever trained by our ConvMix framework outperforms previous baseline methods, which demonstrates our superior effectiveness.

cs.IR↗