Searcharxiv⌕ Search

arXiv subjects

Yasha Wang

Publications and source records attributed to Yasha Wang.

At least 19 recordsLinked to original sources
arXiv2026

AgenticRag-R1: Agentic Reinforcement Learning with Stack Memory for Multi-Step Reasoning, Retrieval and Memorizing

Retrieval-Augmented Generation (RAG) improves the factuality of large language models (LLMs), yet existing RAG systems often struggle with complex, multi-step reasoning that requires adaptive retrieval and continuous revision of intermediate contexts. Recent reinforcement learning (RL)-based agentic RAG methods partially alleviate this issue, but typically rely on coarse-grained action spaces and trajectory-level rewards, resulting in weak reward assignment and a bias toward short-horizon, stereotyped reasoning template. To address, we propose AgenticRag-R1, a RL framework that deeply integrates reasoning, retrieval, and memory via a memory stack and fine-grained action space, supported by hierarchical action-aware rewards and an information-aware trajectory rejection strategy to enable effective long-horizon learning. Experiments across a diverse set of multi-hop, open-domain, and agentic reasoning benchmarks, spanning multiple backbone model sizes, demonstrate that AgenticRag-R1 consistently outperforms strong baselines. Moreover, AgenticRag-R1 learns more robust, interpretable, and memory-aware reasoning behaviors, highlighting the effect of fine-grained action modeling and information-aware optimization for long-horizon reasoning. Our code is anonymous available at https://github.com/jiangxinke/Harness-RL/tree/AgenticRAG-R1-Whitebox.

cs.MA↗
arXiv2026

Harness-RL: Black-Box Reinforcement Learning with Action-Args Decoupling for Central-Agent Multi-Agent Harnesses

Large language model agents increasingly solve long-horizon tasks through multi-agent harnesses in which a central agent coordinates specialized sub-agents, tools, and environments. Training the central policy in such a harness raises two challenges. First, an action label is a low-cardinality decision, whereas its args form a high-dimensional conditional sequence; optimizing both with a shared sequence-level signal can produce conflicting gradients. Second, dynamic scheduling creates interdependent sessions with branches, parallel calls, and rewritten contexts, which cannot be faithfully reduced to one flat token sequence. We introduce Harness-RL, a structured reinforcement learning framework that combines Conflict-Aware Policy Optimization (CAPO) with interface-level black-box trajectory construction. The black-box component captures Interface Call Records, builds per-session prefix trees, and aligns outcome and process rewards with trainable tokens. CAPO uses forward activations to identify parameter partitions associated with action and args tokens, then routes their policy gradients to the corresponding subspaces. Harness-RL supports both central-only and joint multi-agent training. Across seven multi-hop question answering and agentic retrieval benchmarks, it reaches average F1 scores of 42.93 and 47.79 with Qwen2.5-1.5B and Qwen2.5-3B, respectively, while ablations validate the contribution of CAPO and favor central-only optimization in the evaluated setting. Our code is available at https://github.com/jiangxinke/Harness-RL.

cs.MA↗
arXiv2026

LoongReflect: Boosting Long-Horizon Reflection in Search Agents via Global Perspective Distillation

Large language model agents increasingly rely on long-horizon reasoning to solve complex tasks involving planning, tool use, and memory. A critical capability in such settings is reflection: assessing trajectory progress, identifying missing evidence and unreliable intermediate states, and deciding whether to continue, revise, or abandon the current branch. Learning effective reflection, however, is challenging because reflection is performed locally within the current branch, whereas its utility can only be determined by its contribution to the final trajectory outcome. This local-global mismatch makes outcome-based reinforcement learning provide only local, sparse and delayed supervision for reflective decisions. To solve these, we propose LoongReflect, a training framework that formulates reflection as a memory-control policy. The agent operates over a reversible trajectory tree using explicit reflect and backtrack actions. Reflection consolidates verified facts, missing evidence, and branch-specific risks into working memory, while backtracking removes an unreliable branch from the active context and preserves a concise corrective lesson. To learn this policy, LoongReflect combines two complementary signals through a look-ahead, extragradient-style coordination mechanism. A fast channel distills globally informed reflective behavior from a privileged teacher, with supervision restricted to reflection and backtracking tokens. A slow channel optimizes complete trajectories using outcome-based GRPO, aligning local control decisions with final task success. Experiments on multi-hop retrieval-augmented generation and mathematical reasoning benchmarks demonstrate consistent improvements over outcome-only reinforcement learning and self-distillation baselines.

cs.LG↗
arXiv2026

The Weakest Link Tells It All: Outcome-Supervised Process Reward Modeling via Learnable Credit Assignment

Process reward models (PRMs) enhance the reasoning capabilities of large language models (LLMs) by providing fine-grained feedback, yet training PRMs typically requires expensive stepwise annotations. Outcome-supervised PRMs offer a scalable alternative by learning from final-answer correctness alone, but this introduces a fundamental *credit assignment* challenge, i.e., attributing outcomes to responsible reasoning steps. Existing approaches rely on either uniform or causal assignment, both of which fail to anchor credit in step correctness and thus hinder process error identification. In this work, we propose Outcome-Supervised Process Reward Modeling via **L**earnable **C**redit **A**ssignment (**LCA**), an outcome-supervised PRM framework that jointly learns credit assignment and reward modeling under the principle of *Weakest Link Assignment: a reasoning chain is as strong as its weakest link*. To address mutual dependence between credit assignment and reward modeling, we formalize outcome-supervised PRM as a Multiple Instance Learning (MIL) problem and introduce Softmax-Weighted-Sum (SWS) pooling, an MIL pooling technique tailored for strong dependence and redundancy among reasoning states. We prove Bayes consistency of our algorithm under mild assumptions. Extensive experiments demonstrate that **LCA** consistently outperforms state-of-the-art outcome-supervised PRMs across multiple tasks and backbones. Code is available at https://anonymous.4open.science/r/LCA.

cs.LG↗
arXiv2026

Search for Truth from Reasoning: A Dynamic Representation Editing Framework for Steering LLM Trajectories

Current approaches to enhance Large Language Model (LLM) reasoning, such as Chain-of-Thought and "Wait" prompts, primarily encourage models to think more, yet often fail to guide them toward Truth. While Representation Editing (RepE) offers a intrinsic control, its application to dynamic reasoning trajectories remains underexplored. In this work, we bridge this gap by investigating the geometry of truth within unfolding reasoning chains. We uncover three critical insights: (1) Truth is encoded at the sentence level and is entangled with latent reasoning patterns; (2) Effective intervention follows an Uncertainty Principle and a Decay Effect, requiring localization to early, high-entropy forks; (3) Naive steering vectors suffer from noise, risking collateral damage to correct trajectories. Based on these findings, we propose DynaSteer, a dynamic RepE framework. DynaSteer employs pattern clustering to disentangle reasoning manifolds and utilizes Fisher-LDA to project purified truth. By dynamically monitoring lookahead entropy, it selectively steers and rolls back trajectories only when necessary. Comprehensive experimental results on several MATH benchmark verify the effectiveness of DynaSteer, and experiments on out-of-domain coding tasks further confirm its generalization ability. Our code is publicly available at https://github.com/tianlwang/DynaSteer.

cs.AI↗
arXiv2026

Database Context Compression for Text-to-SQL on Real-World Large Databases

Recent progress in Text-to-SQL has been driven by stronger language models and prompting strategies, yet performance on real enterprise benchmarks such as Spider 2.0 and BIRD remains far below that on classical academic datasets. We argue that the main bottleneck is no longer reasoning, but database representation. Real databases contain repeated audit columns, large groups of similar tables, opaque identifiers whose meanings are stored only in documentation, and extensive data dictionaries with little query-relevant information. Existing query-aware methods, including schema linking and retrieval-based schema selection, filter this raw context but still operate on redundant and verbose representations. We reformulate the problem as database context compression, a query-agnostic transformation that rewrites schemas, semantic descriptions, and external documentation into a compact representation. We formalize this transformation with the SGCF (Support-Gain Component Factorization) principle, which unifies repeated column extraction, isomorphic table templating, semantic componentization, and evidence purification under a single coverage objective. Based on SGCF, we propose DBCC, a database-side middleware that performs offline structural and semantic compression together with lightweight online evidence purification. DBCC is model-agnostic and can be integrated into existing Text-to-SQL pipelines. On Spider 2.0-Snow and BIRD, DBCC reduces input context by up to two orders of magnitude (from 2.6M to 34.7K tokens on the largest Spider 2.0-Snow subset), improves schema-linking strict recall from 0% to 56.5% under DeepSeek-V3.2 (63.1% under Claude Opus 4.7), and consistently increases end-to-end execution accuracy by 1.8-1.9% over three recent Text-to-SQL systems. Our code is open-sourced at https://github.com/MrBlankness/SchemaCompression.

cs.DB↗
arXiv2026

EvoRubrics: Dynamic Rubrics as Rewards via Adversarial Co-Evolution for LLM Reinforcement Learning

Rubric-based rewards offer interpretable and fine-grained optimization signals for reinforcement learning in open-ended tasks where verifiable answers are unavailable. However, pre-constructed rubrics remain static throughout training, creating a fundamental mismatch with the evolving policy: fixed criteria gradually lose discriminative power as the model improves, leading to reward saturation and potential hacking. Recent dynamic rubric methods partially address this but rely on external frontier models or ground-truth answers, and update rubrics only at coarse granularity. We propose EvoRubrics, a co-evolutionary RL framework where a Policy LLM and a Rubric Generator jointly improve through adversarial interaction within each training step. As the policy improves under the rubric generator's guidance, the rubric generator adapts its criteria to remain discriminative and informative, enabling evaluation to track the policy in real time and naturally inducing an automatic curriculum. Experiments show that EvoRubrics consistently outperforms static and dynamic rubric baselines across benchmarks. The learned Rubric Generator further generalizes as a transferable reward model. Notably, even a fully self-supervised variant without any external supervision achieves meaningful gains, suggesting that co-evolution between generation and evaluation alone can provide sufficiently rich learning signals. Our code is publicly available at https://anonymous.4open.science/r/EvoRubrics-2155/.

cs.LG↗
arXiv2026

ScaffoldAgent: Utility-Guided Dynamic Outline Optimization for Open-Ended Deep Research

Open-ended deep research (OEDR) requires systems to acquire knowledge through multi-round retrieval and generate coherent long-form reports. The outline plays a central role as a structural scaffold that coordinates retrieval, evidence organization, and generation. However, existing methods either fix the outline before writing or refine it with local heuristics, leading to scaffold drift under continuous information accumulation and delayed feedback for evaluating outline modifications. We propose ScaffoldAgent, a utility-guided dynamic outline optimization framework for OEDR. ScaffoldAgent models outline evolution as a structured decision process with three operations: Expansion, Contraction, and Revision, enabling controlled updates to the report scaffold. It further introduces a utility-guided feedback mechanism that estimates the downstream value of each outline operation from retrieval gain, structural coherence, and trial-generation quality. The resulting utility signal guides node selection, operation scheduling, and termination during inference. Experiments on DeepResearch Bench and DeepResearch Gym show that ScaffoldAgent consistently improves long-form report generation and factual grounding over existing deep research agents.

cs.AI↗
arXiv2026

The Tell-Tale Norm: $\ell_2$ Magnitude as a Signal for Reasoning Dynamics in Large Language Models

Recent work has sought to understand Large Language Models (LLMs) reasoning, yet a principled, model-intrinsic signal that captures its layer-wise reasoning dynamics remains underexplored. We bridge this gap by demonstrating that the l2 norm of hidden states serves as an endogenous signal of the model's reasoning intensity. Using Sparse Autoencoders (SAEs) as a diagnostic probe, we observe that LLMs' internal reasoning is marked by a sharp increase in reasoning feature activations concentrated in late layers. Motivated by this pattern, we establish a formal link between reasoning intensity and the model's latent geometry and theoretically prove that the l2 norm of hidden states bounds the activation strength of SAE reasoning features. Empirical correlation analysis and causal interventions further validate the l2 norm as a faithful indicator, where heightened norms consistently correspond to critical reasoning steps. We then introduce three test-time scaling techniques guided by l2 norms: (i) Adaptive Layer-wise Reasoning Recursion, (ii) Endogenous Reasoning State Steering, and (iii) l2-guided Response Selection, which requires no additional training or data and is compatible with advanced inference engines. Experiments across model architectures and benchmarks show that l2-norm-based techniques significantly improve reasoning performance, offering a principled yet simple lens to perceive and control LLM latent reasoning dynamics. Our code is available at https://github.com/zjy1298/The-Tell-Tale-Norm.

cs.CL↗
arXiv2026

FOREVER: Forgetting Curve-Inspired Memory Replay for Language Model Continual Learning

Continual learning (CL) for large language models (LLMs) aims to enable sequential knowledge acquisition without catastrophic forgetting. Memory replay methods are widely used for their practicality and effectiveness, but most rely on fixed, step-based heuristics that often misalign with the model's actual learning progress, since identical training steps can result in varying degrees of parameter change. Motivated by recent findings that LLM forgetting mirrors the Ebbinghaus human forgetting curve, we propose FOREVER (FORgEtting curVe-inspired mEmory Replay), a novel CL framework that aligns replay schedules with a model-centric notion of time. FOREVER defines model time using the magnitude of optimizer updates, allowing forgetting curve-inspired replay intervals to align with the model's internal evolution rather than raw training steps. Building on this approach, FOREVER incorporates a forgetting curve-based replay scheduler to determine when to replay and an intensity-aware regularization mechanism to adaptively control how to replay. Extensive experiments on three CL benchmarks and models ranging from 0.6B to 13B parameters demonstrate that FOREVER consistently mitigates catastrophic forgetting.

cs.LG↗
arXiv2026

GraphWalker: Patient Analogy Meets Information Gain for Clinical Reasoning with Large Language Models

Clinical reasoning over electronic health records (EHRs) is a fundamental yet challenging task in modern healthcare. While large language models (LLMs) offer a promising paradigm via in-context demonstrations that requires no task-specific parameter updates, existing methods for reasoning by patient analogy in EHR settings suffer from three core limitations: (1) Perspective Limitation, where data-driven similarity misaligns with LLM reasoning needs while model-driven signals are constrained by limited clinical competence; (2) Cohort Awareness, as demonstrations are selected independently without modeling population-level structure; and (3) Information Aggregation, where redundancy and interaction effects among demonstrations are ignored. We propose GraphWalker, a training-free framework that lets frozen LLMs reason by analogy over retrieved patient cases. GraphWalker (i) jointly leverages data-driven and model-driven perspectives, (ii) discovers patient cohorts to ground retrieval in population-level structure, and (iii) employs a lazy greedy search with frontier expansion to compose demonstrations with high marginal information gain. Extensive experiments on multiple real-world EHR benchmarks show that GraphWalker consistently outperforms state-of-the-art demonstration selection baselines, and remains substantially more robust under cross-dataset distribution shift, without task-specific parameter updates. GraphWalker further generalizes to black-box LLMs and composes naturally with agentic reasoning frameworks, positioning it as a pluggable patient-analogy skill in LLM-based clinical workflows. Our code is available at https://github.com/PuppyKnightUniversity/GraphWalker.

cs.LG↗
arXiv2026

Augmenting Clinical Decision-Making with an Interactive and Interpretable AI Copilot: A Real-World User Study with Clinicians in Nephrology and Obstetrics

Clinician skepticism toward opaque AI hinders adoption in high-stakes healthcare. We present AICare, an interactive and interpretable AI copilot for collaborative clinical decision-making. By analyzing longitudinal electronic health records, AICare grounds dynamic risk predictions in scrutable visualizations and LLM-driven diagnostic recommendations. Through a within-subjects counterbalanced study with 16 clinicians across nephrology and obstetrics, we comprehensively evaluated AICare using objective measures (task completion time and error rate), subjective assessments (NASA-TLX, SUS, and confidence ratings), and semi-structured interviews. Our findings indicate AICare's reduced cognitive workload. Beyond performance metrics, qualitative analysis reveals that trust is actively constructed through verification, with interaction strategies diverging by expertise: junior clinicians used the system as cognitive scaffolding to structure their analysis, while experts engaged in adversarial verification to challenge the AI's logic. This work offers design implications for creating AI systems that function as transparent partners, accommodating diverse reasoning styles to augment rather than replace clinical judgment.

cs.HC↗
arXiv2026

StackPlanner: A Centralized Hierarchical Multi-Agent System with Task-Experience Memory Management

Multi-agent systems based on large language models, particularly centralized architectures, have recently shown strong potential for complex and knowledge-intensive tasks. However, central agents often suffer from unstable long-horizon collaboration due to the lack of memory management, leading to context bloat, error accumulation, and poor cross-task generalization. To address both task-level memory inefficiency and the inability to reuse coordination experience, we propose StackPlanner, a hierarchical multi-agent framework with explicit memory control. StackPlanner addresses these challenges by decoupling high-level coordination from subtask execution with active task-level memory control, and by learning to retrieve and exploit reusable coordination experience via structured experience memory and reinforcement learning. Experiments on multiple deep-search and agent system benchmarks demonstrate the effectiveness of our approach in enabling reliable long-horizon multi-agent collaboration.

cs.AI↗
arXiv2025

Bridging Global Intent with Local Details: A Hierarchical Representation Approach for Semantic Validation in Text-to-SQL

Text-to-SQL translates natural language questions into SQL statements grounded in a target database schema. Ensuring the reliability and executability of such systems requires validating generated SQL, but most existing approaches focus only on syntactic correctness, with few addressing semantic validation (detecting misalignments between questions and SQL). As a consequence, effective semantic validation still faces two key challenges: capturing both global user intent and SQL structural details, and constructing high-quality fine-grained sub-SQL annotations. To tackle these, we introduce HEROSQL, a hierarchical SQL representation approach that integrates global intent (via Logical Plans, LPs) and local details (via Abstract Syntax Trees, ASTs). To enable better information propagation, we employ a Nested Message Passing Neural Network (NMPNN) to capture inherent relational information in SQL and aggregate schema-guided semantics across LPs and ASTs. Additionally, to generate high-quality negative samples, we propose an AST-driven sub-SQL augmentation strategy, supporting robust optimization of fine-grained semantic inconsistencies. Extensive experiments conducted on Text-to-SQL validation benchmarks (both in-domain and out-of-domain settings) demonstrate that our approach outperforms existing state-of-the-art methods, achieving an average 9.40% improvement of AUPRC and 12.35% of AUROC in identifying semantic inconsistencies. It excels at detecting fine-grained semantic errors, provides large language models with more granular feedback, and ultimately enhances the reliability and interpretability of data querying platforms.

cs.LG↗
arXiv2025

TPO: Aligning Large Language Models with Multi-branch & Multi-step Preference Trees

In the domain of complex reasoning tasks, such as mathematical reasoning, recent advancements have proposed the use of Direct Preference Optimization (DPO) to suppress output of dispreferred responses, thereby enhancing the long-chain reasoning capabilities of large language models (LLMs). To this end, these studies employed LLMs to generate preference trees via Tree-of-thoughts (ToT) and sample the paired preference responses required by the DPO algorithm. However, the DPO algorithm based on binary preference optimization is unable to learn multiple responses with varying degrees of preference/dispreference that provided by the preference trees, resulting in incomplete preference learning. In this work, we introduce Tree Preference Optimization (TPO), that does not sample paired preference responses from the preference tree; instead, it directly learns from the entire preference tree during the fine-tuning. Specifically, TPO formulates the language model alignment as a Preference List Ranking problem, where the policy can potentially learn more effectively from a ranked preference list of responses given the prompt. In addition, to further assist LLMs in identifying discriminative steps within long-chain reasoning and increase the relative reward margin in the preference list, TPO utilizes Adaptive Step Reward to adjust the reward values of each step in trajectory for performing fine-grained preference optimization. We carry out extensive experiments on mathematical reasoning tasks to evaluate TPO. The experimental results indicate that TPO consistently outperforms DPO across five public large language models on four datasets. Our code is publicly available at https://github.com/MrBlankness/TPO.git.

cs.CL↗
arXiv2025

Magical: Medical Lay Language Generation via Semantic Invariance and Layperson-tailored Adaptation

Medical Lay Language Generation (MLLG) plays a vital role in improving the accessibility of complex scientific content for broader audiences. Recent literature to MLLG commonly employ parameter-efficient fine-tuning methods such as Low-Rank Adaptation (LoRA) to fine-tuning large language models (LLMs) using paired expert-lay language datasets. However, LoRA struggles with the challenges posed by multi-source heterogeneous MLLG datasets. Specifically, through a series of exploratory experiments, we reveal that standard LoRA fail to meet the requirement for semantic fidelity and diverse lay-style generation in MLLG task. To address these limitations, we propose Magical, an asymmetric LoRA architecture tailored for MLLG under heterogeneous data scenarios. Magical employs a shared matrix $A$ for abstractive summarization, along with multiple isolated matrices $B$ for diverse lay-style generation. To preserve semantic fidelity during the lay language generation process, Magical introduces a Semantic Invariance Constraint to mitigate semantic subspace shifts on matrix $A$. Furthermore, to better adapt to diverse lay-style generation, Magical incorporates the Recommendation-guided Switch, an externally interface to prompt the LLM to switch between different matrices $B$. Experimental results on three real-world lay language generation datasets demonstrate that Magical consistently outperforms prompt-based methods, vanilla LoRA, and its recent variants, while also reducing trainable parameters by 31.66%. Our code is publicly available at https://github.com/tianlwang/Magical.git.

cs.CL↗
arXiv2025

DFAMS: Dynamic-flow guided Federated Alignment based Multi-prototype Search

Federated Retrieval (FR) routes queries across multiple external knowledge sources, to mitigate hallucinations of LLMs, when necessary external knowledge is distributed. However, existing methods struggle to retrieve high-quality and relevant documents for ambiguous queries, especially in cross-domain scenarios, which significantly limits their effectiveness in supporting downstream generation tasks. Inspired by Dynamic Information Flow (DIF), we propose DFAMS, a novel framework that leverages DIF to identify latent query intents and construct semantically aligned knowledge partitions for accurate retrieval across heterogeneous sources. Specifically, DFAMS probes the DIF in LLMs by leveraging gradient signals from a few annotated queries and employing Shapley value-based attribution to trace neuron activation paths associated with intent recognition and subdomain boundary detection. Then, DFAMS leverages DIF to train an alignment module via multi-prototype contrastive learning, enabling fine-grained intra-source modeling and inter-source semantic alignment across knowledge bases. Experimental results across five benchmarks show that DFAMS outperforms advanced FR methods by up to 14.37\% in knowledge classification accuracy, 5.38\% in retrieval recall, and 6.45\% in downstream QA accuracy, demonstrating its effectiveness in complex FR scenarios. Our code are anonymous available at https://anonymous.4open.science/r/DFAMS/

cs.LG↗
arXiv2025

ADEPT: Continual Pretraining via Adaptive Expansion and Dynamic Decoupled Tuning

Conventional continual pretraining (CPT) for large language model (LLM) domain adaptation often suffers from catastrophic forgetting and limited domain capacity. Existing strategies adopt layer expansion, introducing additional trainable parameters to accommodate new knowledge. However, the uniform expansion and updates still entangle general and domain learning, undermining its effectiveness. Our pilot studies reveal that LLMs exhibit functional specialization, where layers and units differentially encode general-critical capabilities, suggesting that parameter expansion and optimization should be function-aware. We then propose ADEPT, Adaptive Expansion and Dynamic Decoupled Tuning for continual pretraining, a two-stage framework for domain-adaptive CPT. ADEPT first performs General-Competence Guided Selective Layer Expansion, duplicating layers least critical for the general domain to increase representational capacity while minimizing interference with general knowledge. It then applies Adaptive Unit-Wise Decoupled Tuning, disentangling parameter units within expanded layers according to their general-domain importance and assigning asymmetric learning rates to balance knowledge injection and retention. Experiments on mathematical and medical benchmarks show that ADEPT outperforms full-parameter CPT by up to 5.76% on the general domain and 5.58% on the target domain with only 15% of parameters tuned and less than 50% training time. Ablation studies, theoretical analysis, and extended investigations further demonstrate the necessity of targeted expansion and decoupled optimization, providing new principles for efficient and robust domain-adaptive CPT. Our code is open-sourced at https://github.com/PuppyKnightUniversity/ADEPT

cs.LG↗