arXiv subjects

Shuangyong Song

Publications and source records attributed to Shuangyong Song.

At least 19 recordsLinked to original sources

arXiv2026

Quality Determines Direction, Length Shapes Magnitude: Length Control for Open-Ended Reinforcement Learning

Reinforcement learning (RL) changes not only what language models say, but also how much they say, often increasing response length at the cost of token efficiency. Controlling this length growth is particularly challenging in open-ended RL because (i) response length is entangled with quality, (ii) open-ended tasks lack a natural success boundary for deciding when efficiency should be prioritized, and (iii) dense, graded rewards often yield small within-group quality margins, making quality-induced advantages especially sensitive to reward-level length shaping, which can perturb their magnitudes and even reverse their signs. We therefore adopt an asymmetric principle: quality should determine the direction of reinforcement, while length should only shape its magnitude. We instantiate this principle with Quality-Gated Length Advantage Shaping (QGLAS), which first computes advantages from quality rewards alone, then adds bounded bonuses only to shorter positive-advantage responses, leaving all other advantages unchanged. The bonus strength is further adapted to within-group quality separation, allowing conciseness to matter more when quality-favored responses are similar and less when their quality differences are clear. Across different model families, open-ended benchmarks, and reward sources, QGLAS consistently achieves a stronger quality--length trade-off than representative baselines. At approximately 30% compression, QGLAS retains 98.4--102.0% of the macro-average quality gains achieved by quality-only RL over the base model, compared with 68.3--75.5% for these baselines at comparable compression.

cs.LG↗

arXiv2026

Dual Attention Residuals

Recent work extends Transformer residual pathways along two complementary axes: historical retrieval selects information from earlier depths, whereas multi-stream methods maintain multiple residual trajectories. These capabilities have largely been studied in isolation, and assigning an independent retriever to each stream still prevents one trajectory from influencing depth selection in another. We propose Dual Attention Residuals (DAR), which brings multi-stream interaction into historical retrieval through reciprocal cross-stream addressing. For each target stream, DAR computes depth weights from normalized states in the opposite stream and applies them to values from the target stream's own history. The retrieved states are combined for an unchanged Transformer branch and updated through constrained gated writes; a block-form variant operates on block-level histories to control overhead. Across dense models from 0.1B to 1B parameters and a 7B sparse-MoE model, DAR consistently improves validation loss over standard residual Transformers and Attention Residuals. Routing ablations show that the gain cannot be explained by an additional stream or value projection alone. Representation and intervention analyses further show that reciprocal cross-stream selection preserves depth-wise diversity and avoids the redundancy or functional imbalance observed in alternative two-stream designs.

cs.CL↗

arXiv2026

FCPRAG: Fusion-Controller Parametric Retrieval-Augmented Generation for Stable Multi-Passage LoRA Injection

Parametric retrieval-augmented generation (PRAG) injects retrieved evidence into a large language model (LLM) through passage-specific LoRA adapters, reducing reliance on long in-context prompts. When multiple passages are retrieved for the same query, however, evidence-level fusion becomes a bottleneck: equal-weight merging can amplify weak or conflicting evidence, and translating retrieval signals into fusion weights often requires fragile global tuning. We propose FCPRAG, a fusion-controlled parametric RAG framework that adds a lightweight controller for retrieval-conditioned, sample-level adapter fusion. The controller predicts per-passage fusion scores together with sample-level calibration signals, including a mixing gate and an adaptive temperature, enabling fusion that stays selective under informative retrieval signals and conservative under uncertainty. FCPRAG is trained with merge-aware supervision derived from each adapter's marginal contribution within a multi-adapter merge, using training data only. We further show that a single dataset-level temperature is suboptimal under heteroscedastic retrieval uncertainty, motivating sample-level adaptation. Experiments on HotpotQA, 2WikiMultiHopQA, PopQA, and ComplexWebQuestions (CWQ) across three LLM backbones show that FCPRAG consistently improves F1 over standard RAG and parametric RAG baselines, with gains of up to 4.65% on 2WikiMultiHopQA and 7.55% on CWQ, while also reducing tuning cost and improving robustness under retrieval perturbations.

cs.CL↗

arXiv2026

Learning When Not to Listen: Selective Anti-Interference Pretraining for Language Models

Language models can over-condition on irrelevant preceding text: predictions already supported by local context may still change when distant, unrelated prefix tokens are perturbed. This interference is especially consequential in long, packed, or distractor-heavy contexts, where useful evidence and irrelevant spans coexist. We propose Selective Prefix Anti-Interference Regularization (SPAR), a pretraining objective for selective anti-interference. SPAR runs the original sequence and a corrupt-prefix input in which only the far prefix is changed, then uses a short-context sufficiency gate and a gated KL objective to stabilize locally supported suffix predictions. The gate operationalizes a model-based estimate of whether the far prefix supplies additional information about the target token. Mechanism analyses show that the gate identifies locally sufficient tokens and sharply reduces prefix sensitivity on gate-selected suffix tokens. In continued training on pretrained base models, SPAR improves RULER across Qwen2.5-0.5B, Qwen2.5-3B, Llama-3.2-1B, Llama-3.1-8B, and GPT2-XL under equal counted training compute; pretraining experiments further show gains on both RULER and NoLiMa. These results show that selective anti-interference is an effective objective-level signal for robust context use.

cs.CL↗

arXiv2026

Multi-Objective Exploration and Preference Optimization via Mutual Information

Aligning large language models with diverse and heterogeneous human values requires multi-objective alignment methods to effectively trade off conflicting preference dimensions. Current methods achieve this trade-off by training policies conditioned on preference vectors and leveraging online direct preference optimization. However, exploration uncertainty can cause the reward distributions of responses generated under different preference vectors to overlap, and the generated responses may fail to effectively align with the corresponding preference vectors. In this paper, we propose Multi-Objective Exploration and Preference Optimization via Mutual Information (MI-EPO), an information-theoretic framework. It unifies multi-objective exploration and alignment by maximizing the joint conditional mutual information among generated responses, preference feedback, and preference vectors. By incorporating a probabilistic routing mechanism, MI-EPO naturally decomposes objective alignment and preference-aware exploration, encouraging the model to generate responses that are distinguishable and aligned with different preference conditions. Experiments on safe alignment and helpful assistant tasks show that MI-EPO significantly improves the alignment between generated responses and preference vectors, makes the outputs more controllable, and achieves stable trade-offs across multiple objectives.

cs.CL↗

arXiv2026

CRAFT: A Unified Counterfactual Reasoning Framework for Tabular Question Answering and Fact Verification

Table reasoning remains challenging for large language models (LLMs), particularly in tasks that require multi-step inference over long and structured tables. Existing approaches predominantly rely on single-direction reasoning, which limits their ability to explore alternative hypotheses across tasks. In this work, we propose CRAFT, a unified Counterfactual Reasoning Framework that reformulates Tabular question answering and fact verification into a general bidirectional verification process. Our method explicitly constructs both declarative statements and their counterfactual variants. Evidence is then extracted from reasoning along both the original and counterfactual paths, and integrated via a weighted mechanism to arrive at the final answer. Experimental results show that our approach consistently surpasses representative baselines on table reasoning datasets such as WikiTQ and TabFact, achieving especially large improvements on complex question answering. Our framework also significantly mitigates performance gaps between different backbone LLMs. This indicates that counterfactual reasoning effectively overcomes the limitations of single-direction inference, guiding LLMs toward more discerning reasoning and establishing a more principled paradigm for structured reasoning tasks. Our code will be made publicly available upon acceptance.

cs.CL↗

arXiv2026

Hallucinations as Orthogonal Noise: Inference-Time Manifold Alignment via Dynamic Contextual Orthogonalization

Hallucination in Large Language Models (LLMs), characterized by the generation of content inconsistent with contextual facts or logical constraints -- remains a persistent challenge for reliable deployment. In this work, we address this issue through a geometric framework rooted in the linear representation hypothesis. We propose that hallucinations manifest as orthogonal noise relative to the semantic manifold of the residual stream. Specifically, we hypothesize that while attention heads ideally propagate information congruent with the context subspace, hallucinations arise when specific heads introduce components orthogonal to this subspace, disrupting the coherence of the latent representation. Based on this formulation, we introduce Dynamic Contextual Orthogonalization (DCO), an inference-time intervention method. DCO utilizes the input residual stream as a dynamic context anchor to perform orthogonal decomposition on attention head outputs. To distinguish between context-aligned semantic updates and divergent noise, DCO employs a layer-wise Z-score suppression mechanism that selectively attenuates outlier orthogonal components based on statistical distributions. Evaluations on Llama-3-8B and 70B across benchmarks such as XSum, NQ-Swap, and IFEval demonstrate that DCO achieves superior contextual faithfulness compared to state-of-the-art intervention baselines. Furthermore, DCO maintains high performance on knowledge-intensive tasks like TriviaQA and TruthfulQA, effectively mitigating the trade-off between hallucination suppression and parametric knowledge retention often observed in existing methods. Our findings validate the geometric interpretation of hallucinations and establish DCO as a computationally efficient approach for enforcing manifold alignment.Our code is available at https://github.com/Harry-Miral/DCO

cs.CL↗

arXiv2026

Prompt-Level Reward Specifications for Open-Ended Post-Training

Open-ended post-training benefits from rewards that make prompt-specific success conditions explicit, rather than relying only on post-hoc scalar scores. In instruction following, writing, and decision-support tasks, response quality depends on local requirements, holistic preferences, and explicit constraints, but existing reward methods often leave these criteria implicit or cover only narrowly verifiable cases. We propose a prompt-level reward specification framework that separates reward specification from reward computation. Given only prompts, our framework constructs reusable task-adaptive rubrics and executable hard-constraint checkers offline, making reward criteria explicit before training and reusable across rollouts. At scoring time, artifact-anchored rubric and code scores are combined with an independent global score for residual holistic quality, yielding a normalized hybrid reward over requirement satisfaction, holistic quality, and deterministic constraints. The framework requires no human preference annotations, reference answers, or a separately trained reward model. Experiments show that the resulting reward improves offline RM-style response ranking and supports online reinforcement learning across multiple open-ended benchmarks. Ablations further show that rubrics, global scoring, and executable verification provide complementary supervision.

cs.CL↗

arXiv2026

Stop Rewarding Hallucinated Steps: Faithfulness-Aware Step-Level Reinforcement Learning for Small Reasoning Models

As large language models become smaller and more efficient, small reasoning models (SRMs) are crucial for enabling chain-of-thought (CoT) reasoning in resource-constrained settings. However, they are prone to faithfulness hallucinations, especially in intermediate reasoning steps. Existing mitigation methods based on online reinforcement learning rely on outcome-based rewards or coarse-grained CoT evaluation, which can inadvertently reinforce unfaithful reasoning when the final answer is correct. To address these limitations, we propose Faithfulness-Aware Step-Level Reinforcement Learning (FaithRL), introducing step-level supervision via explicit faithfulness rewards from a process reward model, together with an implicit truncated resampling strategy that generates contrastive signals from faithful prefixes, while also mitigating reward hacking from step-level rewards. Experiments across multiple SRMs and Open-Book QA benchmarks demonstrate that FaithRL consistently reduces hallucinations in both the CoT and final answers, leading to more faithful and reliable reasoning. Code is available at https://github.com/Easy195/FaithRL.

cs.CL↗

arXiv2026

Pairwise Preference Reward and Group-Based Diversity Enhancement for Superior Open-Ended Generation

Current reinforcement learning(RL) methods are broadly applicable and powerful in verifiable settings where scalar rewards can be provided. However, in open-ended generation tasks, verifying the correctness of responses remains challenging, and training reward models incurs substantial computational and annotation costs. Moreover, reinforcement learning (RLVR) often leads to diversity collapse and produces stereotypical or rigid outputs, outcomes that are particularly undesirable in open-domain scenarios. We propose Pairwise Preference Reward and Group-based Diversity Enhancement (PPR-GDE), a RL method that is more suitable for open-ended generation. PPR-GDE does not require scalar rewards and incorporates group-level diversity into the reward signal, it preserves the comparative structure of subjective evaluation through a pairwise preference reward, mitigates judge position bias via repeated comparisons with swapped response order, and introduces a group-based diversity reward that explicitly encourages semantic dispersion within a response group, all of these reward signals are integrated into a unified group-relative policy optimization objective. We instantiate PPR-GDE on role-playing task, experiments show that PPR-GDE achieves a better alignment quality as well as expressive diversity than strong RL baselines. Further analysis shows that pairwise preference is critical for preference alignment in subjective perspective, while the diversity metric plays an essential role in achieving superior expressive diversity and broader semantic coverage.

cs.AI↗

arXiv2026

Table-R1: Region-based Reinforcement Learning for Table Understanding

Tables present unique challenges for language models due to their structured row-column interactions, necessitating specialized approaches for effective comprehension. While large language models (LLMs) have demonstrated potential in table reasoning through prompting and techniques like chain-of-thought (CoT) and program-of-thought (PoT), optimizing their performance for table question answering remains underexplored. In this paper, we introduce region-based Table-R1, a novel reinforcement learning approach that enhances LLM table understanding by integrating region evidence into reasoning steps. Our method employs Region-Enhanced Supervised Fine-Tuning (RE-SFT) to guide models in identifying relevant table regions before generating answers, incorporating textual, symbolic, and program-based reasoning. Additionally, Table-Aware Group Relative Policy Optimization (TARPO) introduces a mixed reward system to dynamically balance region accuracy and answer correctness, with decaying region rewards and consistency penalties to align reasoning steps. Experiments show that Table-R1 achieves an average performance improvement of 14.36 points across multiple base models on three benchmark datasets, even outperforming baseline models with ten times the parameters, while TARPO reduces response token consumption by 67.5% compared to GRPO, significantly advancing LLM capabilities in efficient tabular reasoning.

cs.CL↗

arXiv2026

Awakening Dormant Experts:Counterfactual Routing to Mitigate MoE Hallucinations

Sparse Mixture-of-Experts (MoE) models have achieved remarkable scalability, yet they remain vulnerable to hallucinations, particularly when processing long-tail knowledge. We identify that this fragility stems from static Top-$k$ routing: routers tend to favor high-frequency patterns over rare factual associations. Consequently, ``specialist experts'' possessing critical long-tail knowledge are often assigned low gating scores and remain ``dormant'' -- under-prioritized for specific tokens despite their proven causal importance on other inputs. To address this, we propose Counterfactual Routing (CoR), a training-free inference framework designed to awaken these dormant experts. CoR integrates layer-wise perturbation analysis with the Counterfactual Expert Impact (CEI) metric to dynamically shift computational resources from syntax-dominant to knowledge-intensive layers while maintaining a constant total activation count, effectively retrieving causally decisive experts via virtual ablation. Extensive experiments on TruthfulQA, FACTOR, and TriviaQA demonstrate that CoR improves factual accuracy by 3.1\% on average without increasing the inference budget, establishing a superior Pareto frontier compared to static scaling strategies.

cs.LG↗

arXiv2026

D-QRELO: Training- and Data-Free Delta Compression for Large Language Models via Quantization and Residual Low-Rank Approximation

Supervised Fine-Tuning (SFT) accelerates taskspecific large language models (LLMs) development, but the resulting proliferation of finetuned models incurs substantial memory overhead. Delta compression addresses this by retaining a single pre-trained LLM with multiple compressed delta weights. However, existing methods fail on models fine-tuned with largescale datasets. We find that larger SFT data scale amplifies delta parameter magnitude, singular values, and entropy, exacerbating compression errors. To tackle this, we propose DQRELO (Delta Compression via Quantization and Residual Low-Rank), a novel training- and data-free delta compression method. It combines coarse-grained one-bit quantization to capture the dominant structure of the delta, followed by compensated residual low-rank approximation to recover fine-grained details from the smaller residual error. Experiments on various LLMs spanning dense and MoE architectures across multiple domains under this challenging setting demonstrate that DQRELO outperforms existing methods. Moreover, we establish key design principles for delta compression through extensive empirical analysis, demonstrating how task difficulty, architecture, and layer positioning create predictable patterns that can guide optimal compression strategies in production systems.

cs.LG↗

arXiv2026

Dynamic Knowledge Fusion for Multi-Domain Dialogue State Tracking

The performance of task-oriented dialogue models is strongly tied to how well they track dialogue states, which records and updates user information across multi-turn interactions. However, current multi-domain DST encounters two key challenges: the difficulty of effectively modeling dialogue history and the limited availability of annotated data, both of which hinder model performance. To tackle the aforementioned problems, we develop a dynamic knowledge fusion framework applicable to multi-domain DST. The model operates in two stages: first, an encoder-only network trained with contrastive learning encodes dialogue history and candidate slots, selecting relevant slots based on correlation scores; second, dynamic knowledge fusion leverages the structured information of selected slots as contextual prompts to enhance the accuracy and consistency of dialogue state tracking. This design enables more accurate integration of dialogue context and domain knowledge. Results obtained from multi-domain dialogue benchmarks indicate that our method notably improves both tracking accuracy and generalization, validating its capability in handling complex dialogue scenarios.

cs.CL↗

arXiv2026

Chain-of-Lure: A Universal Jailbreak Attack Framework using Unconstrained Synthetic Narratives

In the era of rapid generative AI development, interactions with large language models (LLMs) pose increasing risks of misuse. Prior research has primarily focused on attacks using template-based prompts and optimization-oriented methods, while overlooking the fact that LLMs possess strong unconstrained deceptive capabilities to attack other LLMs. This paper introduces a novel jailbreaking method inspired by the Chain-of-Thought mechanism. The attacker employs mission transfer to conceal harmful user intent within dialogue and generates a progressive chain of lure questions without relying on predefined templates, enabling successful jailbreaks. To further improve the attack's strength, we incorporate a helper LLM model that performs randomized narrative optimization over multi-turn interactions, enhancing the attack performance while preserving alignment with the original intent. We also propose a toxicity-based framework using third-party LLMs to evaluate harmful content and its alignment with malicious intent. Extensive experiments demonstrate that our method consistently achieves high attack success rates and elevated toxicity scores across diverse types of LLMs under black-box API settings. These findings reveal the intrinsic potential of LLMs to perform unrestricted attacks in the absence of robust alignment constraints. Our approach offers data-driven insights to inform the design of future alignment mechanisms. Finally, we propose two concrete defense strategies to support the development of safer generative models. Our code is available at https://github.com/ChangWenhan/chain-of-lure-official

cs.CR↗

arXiv2026

UniARM: Towards a Unified Autoregressive Reward Model for Multi-Objective Test-Time Alignment

Multi-objective alignment aims to align LLM responses with multiple human preference objectives. Among existing methods, guiding the generation of frozen LLMs through autoregressive reward models (ARMs) to accomplish multi-objective test-time alignment is a low-cost solution. However, these methods typically rely on independent parameters for each preference objective, either by training ARMs independently across preference dimensions, which neglects interactions among preference features, or by training a single ARM with separate feature extraction modules for each preference, which can cause feature entanglement. Both strategies can result in misalignment between generated outputs and user preferences. To address this limitation, we propose Preference-Modulated \& Shared Low-Rank Adaptation (MoSLoRA) for ARM training, which first extracts shared features via a preference-agnostic module and then applies affine transformations to shared features via a preference modulation module conditioned on mixed preference vectors. This design mitigates feature entanglement and enables precise control over preference trade-offs during inference. Building on this, we introduce the Unified Autoregressive Reward Model (UniARM), a novel framework for multi-objective test-time alignment. UniARM jointly models all preference dimensions in a single parameter space, eliminating the need for independent parameters for each preference objective. es on larger-scale LLMs, enhancing its practical usability.

cs.CL↗

arXiv2026

A benchmark for joint dialogue satisfaction, emotion recognition, and emotion state transition prediction

User satisfaction is closely related to enterprises, as it not only directly reflects users' subjective evaluation of service quality or products, but also affects customer loyalty and long-term business revenue. Monitoring and understanding user emotions during interactions helps predict and improve satisfaction. However, relevant Chinese datasets are limited, and user emotions are dynamic; relying on single-turn dialogue cannot fully track emotional changes across multiple turns, which may affect satisfaction prediction. To address this, we constructed a multi-task, multi-label Chinese dialogue dataset that supports satisfaction recognition, as well as emotion recognition and emotional state transition prediction, providing new resources for studying emotion and satisfaction in dialogue systems.

cs.CL↗

arXiv2026

ReasonTabQA: A Comprehensive Benchmark for Table Question Answering from Real World Industrial Scenarios

Recent advancements in Large Language Models (LLMs) have significantly catalyzed table-based question answering (TableQA). However, existing TableQA benchmarks often overlook the intricacies of industrial scenarios, which are characterized by multi-table structures, nested headers, and massive scales. These environments demand robust table reasoning through deep structured inference, presenting a significant challenge that remains inadequately addressed by current methodologies. To bridge this gap, we present ReasonTabQA, a large-scale bilingual benchmark encompassing 1,932 tables across 30 industry domains such as energy and automotive. ReasonTabQA provides high-quality annotations for both final answers and explicit reasoning chains, supporting both thinking and no-thinking paradigms. Furthermore, we introduce TabCodeRL, a reinforcement learning method that leverages table-aware verifiable rewards to guide the generation of logical reasoning paths. Extensive experiments on ReasonTabQA and 4 TableQA datasets demonstrate that while TabCodeRL yields substantial performance gains on open-source LLMs, the persistent performance gap on ReasonTabQA underscores the inherent complexity of real-world industrial TableQA.

cs.CL↗