AAAI 2026 Papers

Reward models (RMs) are a core component in the post-training of large language models (LLMs), serving as proxies for human preference evaluation and guiding model alignment. However, training reliable RMs under limited resources remains challenging due to the reliance on large-scale preference annotations and the high cost of fine-tuning LLMs. To address this, we propose SparseRM, which leverages Sparse Autoencoder (SAE) to extract preference-relevant information encoded in model representations, enabling the construction of a lightweight and interpretable reward model. SparseRM first employs SAE to decompose LLM representations into interpretable directions that capture preference-relevant features. The representations are then projected onto these directions to compute alignment scores, which quantify the strength of each preference feature in the representations. A simple reward head aggregates these scores to predict preference scores. Experiments on three preference modeling tasks show that SparseRM achieves superior performance over most mainstream RMs while using less than 1% of trainable parameters. Moreover, it integrates seamlessly into downstream alignment pipelines, highlighting its potential for efficient alignment.

Tree search-based methods have made significant progress in enhancing the code generation capabilities of large language models. However, due to the difficulty in effectively evaluating intermediate algorithmic steps and the inability to locate and timely correct erroneous steps, these methods often generate incorrect code and incur increased computational costs. To tackle these problems, we propose RPM-MCTS, an effective method that utilizes Knowledge-Retrieval as Process Reward Model based on Monte Carlo Tree Search to evaluate intermediate algorithmic steps. By utilizing knowledge base retrieval, RPM-MCTS avoids the complex training of process reward models. During the expansion phase, similarity filtering is employed to remove redundant nodes, ensuring diversity in reasoning paths. Furthermore, our method utilizes sandbox execution feedback to locate erroneous algorithmic steps during generation, enabling timely and targeted corrections. Extensive experiments on four public code generation benchmarks demonstrate that RPM-MCTS outperforms current state-of-the-art methods while achieving an approximately 15% reduction in token consumption. Furthermore, full fine-tuning of the base model using the data constructed by RPM-MCTS significantly enhances its code capabilities.

With the rapid deployment of Chinese large language models (LLMs), culturally-grounded bias evaluation remains understudied due to the dominance of English benchmarks and simplistic Chinese scenarios. To address this, we propose GeWu, a comprehensive benchmark featuring a culturally-aware dataset of 60,192 questions spanning 14 social groups with fine-grained Chinese contexts, significantly exceeding existing resources in breadth and depth. Our two-stage evaluation first quantifies bias via multiple-choice questions using a novel probability-based scoring mechanism to sensitively capture bias tendencies, distilling high-bias scenarios into GeWu-1K. This refined subset then enables multi-turn dialogue evaluations for in-depth analysis under realistic conditions. Experiments reveal that GeWu effectively exposes social biases in state-of-the-art Chinese LLMs, with 13.93% of scenarios eliciting universal bias across all models. This highlights persistent challenges and provides actionable insights for bias mitigation in Chinese contexts.

Weakly supervised phrase localization (WSPL) aims to localize visual objects mentioned by given phrases, but it learns without human-annotated bounding boxes. Previous works struggle in multi-object scenarios where objects in the background often appear simultaneously with the target objects. To this end, we propose a Diffusion-Assisted PrOgressive learning framework (i.e., DAPO) for WSPL task in this paper. Specifically, we score the difficulty of training samples based on the quantity of objects and the level of semantic alignment. These samples are then used progressively during training, in an order by their difficulty scores. To address the sample imbalance problem, we propose a Generation-Assisted Tuning (GAT) method for the grounding network. First, to enrich the samples from few-object scenarios, we leverage Stable Diffusion (SD) to generate images with phrases. Second, we introduce an attention-driven scheme to direct SD's attention on the mentioned objects. Finally, we design a diffusion-guided loss, which helps the grounding network learn the objects' layouts. Extensive experiments show that our DAPO framework outperforms the strong baselines on benchmark datasets.

As Audio Large Language Models (ALLMs) emerge as powerful tools for speech processing, their safety implications demand urgent attention. While considerable research has explored textual and vision safety, audio’s distinct characteristics present significant challenges. This paper first investigates: Is ALLM vulnerable to backdoor attacks exploiting acoustic triggers? In response to this issue, we introduce Hidden in the Noise (HIN), a novel backdoor attack framework designed to exploit subtle, audio-specific features. HIN applies acoustic modifications to raw audio waveforms, such as alterations to temporal dynamics and strategic injection of spectrally tailored noise. These changes introduce consistent patterns that an ALLM’s acoustic feature encoder captures, embedding robust triggers within the audio stream. To evaluate ALLM robustness against audio-feature-based triggers, we develop the AudioSafe benchmark, assessing nine distinct risk types. Extensive experiments on AudioSafe and three established safety datasets reveal critical vulnerabilities in existing ALLMs: (I) audio features like environment noise and speech rate variations achieve over 90% average attack success rate, (II) ALLMs exhibit significant sensitivity differences across acoustic features, particularly showing minimal response to volume as a trigger, and (III) poisoned sample inclusion causes only marginal loss curve fluctuations, highlighting the attack’s stealth.

Recent advances in large-scale code generation models have led to remarkable progress in producing high-quality code. These models are trained in a self-supervised manner on extensive unlabeled code corpora using a decoder-only architecture. However, despite their generative strength, decoder-only models often exhibit limited performance on code understanding tasks such as code search and clone detection, primarily due to their generation-oriented training objectives. While training large encoder-only models from scratch on massive code datasets can improve understanding ability but remains computationally expensive and time-consuming. In this paper, we explore a more efficient alternative by transferring knowledge from pre-trained decoder-only code generation models to code understanding tasks. We investigate how decoder-only architectures can be effectively adapted to learn discriminative and semantically meaningful code representations. To this end, we propose CL4D, a contrastive learning framework tailored to strengthen the representation capabilities of decoder-only models. Extensive experiments on multiple benchmark datasets demonstrate that CL4D achieves competitive or superior performance compared to existing methods on representative code understanding tasks, including code search and clone detection. Further analysis reveals that CL4D substantially improves the semantic alignment of code representations by reducing the distance between semantically similar code snippets. These findings highlight the feasibility of leveraging decoder-only models as a unified backbone for both code generation and understanding.

Knowledge distillation (KD) is a promising compression technique for reducing the computational burden of large language models (LLMs). Depending on access to the teacher model’s internal parameters, KD is typically categorized into white-box and black-box KD. While white-box KD benefits from full access to intrinsic knowledge such as softmax distributions, black-box KD adopts a black-box LLM (e.g., GPT-4) as the teacher, which provides only text-level outputs via API calls. This limited supervision makes black-box KD generally less effective than its white-box counterpart. To bridge the gap between white-box and black-box KD, we propose GrayKD, a novel framework that can effectively distill text-level knowledge from a black-box LLM in a single-stage manner. In particular, rationales generated by the black-box LLM are injected into the student via a lightweight cross-attention module (teacher mode), enabling the model to approximate the black-box teacher’s output distribution without access to internal parameters. The student is then trained with the softmax-level knowledge provided by the teacher mode (student mode). Since both the teacher and student modes share the same backbone, the proposed teacher mode remains highly parameter-efficient, requiring only a small number of additional parameters for rationale injection. Experimental results on instruction-following tasks demonstrate that GrayKD achieves substantial performance improvements over existing KD methods.

Self-play fine-tuning has emerged as a promising approach to improve Large Language Models (LLMs) without additional human annotations. However, existing methods struggle with complex generation tasks requiring long context understanding, where models produce partially correct outputs interleaved with errors. Traditional approaches train on entire sequences uniformly, failing to distinguish between well-predicted and erroneous regions, leading to diluted learning signals and slow convergence. We propose DRIFT (Difference-aware Reinforcement through Iterative Fine-Tuning), a novel self-play framework that selectively trains on prediction differences. DRIFT introduces two key innovations: (1) Difference-Aware Masking (DAM) that identifies and masks common subsequences between model outputs and ground truth, focusing training exclusively on error regions; (2) Occurrence-Aware Loss (OAL) that provides position-invariant vocabulary supervision, complementing the position-sensitive adversarial loss. This dual mechanism enables models to correct both positional and lexical errors effectively. Theoretically, we prove that DRIFT converges when masked distributions align. Empirically, we evaluate DRIFT on diverse summarization benchmarks using Qwen2.5-3B and LLaMA-3.1-8B models. Results show that DRIFT significantly outperforms both supervised fine-tuning (SFT) and self-play fine-tuning (SPIN), achieving up to 16\% improvement on SAMSum dialogue summarization tasks while maintaining general capabilities. Notably, DRIFT breaks the performance ceiling of continued SFT and demonstrates superior efficiency compared to holistic self-play methods, validating that targeted optimization on prediction differences is crucial for structured text generation tasks.

Hallucination has emerged as a pivotal challenge of Large Language Models (LLMs) that generate plausible yet non‑factual content, significantly impeding the trustworthy AI applications in real-world scenarios like medical diagnosis and autonomous driving. Editing the internal activations of LLMs during inference has shown promising effectiveness in mitigating hallucinations with minimal cost. However, previous editing approaches neglect the query‑specific inference pathways that require tailored truthful steering vectors, resulting in suboptimal hallucination mitigation. To address these issues, we propose the Query-Routed Activation Editing (QRAE) framework, which comprises Divergence-sensitive Head Routing (DHR) and Truth-hierarchical Preference Steering (TPS), to fully leverage query-specific semantics for adaptive activation editing. Specifically, DHR is proposed to establish a query-aware head selection criterion, thereby dynamically routing to truth-critical attention heads. Subsequently, TPS introduces a query-specific steering vector calibration policy with the guidance of progressive truth-preferred optimization, enabling precise and adaptive editing for each distinct query. Extensive experiments on the widely recognized TruthfulQA benchmark demonstrate that QRAE outperforms SOTA methods by up to 13.2% in MC1. Meanwhile, QRAE demonstrates strong generalization to out-of-distribution TriviaQA and Natural Questions benchmarks.

Large language models are extensively utilized in creative writing applications. Creative writing requires a balance between subjective writing quality (e.g., literariness and emotional expression) and objective constraint following (e.g., format requirements and word limits). Existing reinforcement learning methods struggle to balance these two aspects: single reward strategies fail to improve both abilities simultaneously, while fixed-weight mixed-reward methods lack the ability to adapt to different writing scenarios. To address this problem, we propose Reinforcement Learning with Mixed Rewards (RLMR), utilizing a dynamically mixed reward system from a writing reward model evaluating subjective writing quality and a constraint verification model assessing objective constraint following. The constraint following reward weight is adjusted dynamically according to the writing quality within sampled groups, ensuring that samples violating constraints get negative advantage in GRPO and thus penalized during training, which is the key innovation of this proposed method. We conduct automated and manual evaluations across diverse model families from 8B to 72B parameters. Additionally, we construct a real-world writing benchmark named WriteEval for comprehensive evaluation. Results illustrate that our method achieves consistent improvements in both instruction following (IFEval from 83.36% to 86.65%) and writing quality (72.75% win rate in manual expert pairwise evaluations on WriteEval). To the best of our knowledge, RLMR is the first work to combine subjective preferences with objective verification in online RL training, providing an effective solution for multi-dimensional creative writing optimization.

Long-context inference in large language models (LLMs) is increasingly constrained by the KV cache bottleneck: memory usage grows linearly with sequence length, while attention computation scales quadratically. Existing approaches address this issue by compressing the KV cache along the temporal axis through strategies such as token eviction or merging to reduce memory and computational overhead. However, these methods often neglect fine-grained importance variations across feature dimensions (i.e., the channel axis), thereby limiting their ability to effectively balance efficiency and model accuracy. In reality, we observe that channel saliency varies dramatically across both queries and positions: certain feature channels carry near-zero information for a given query, while others spike in relevance. To address this oversight, we propose SPARK, a training-free plug-and-play method that applies unstructured sparsity by pruning KV at the channel level, while dynamically restoring the pruned entries during attention score computation. Notably, our approach is orthogonal to existing KV compression and quantization techniques, making it compatible for integration with them to achieve further acceleration. By reducing channel-level redundancy, SPARK enables processing of longer sequences within the same memory budget. For sequences of equal length, SPARK not only preserves or improves model accuracy but also reduces KV cache storage by over 30% compared to eviction-based methods. Furthermore, even in an aggressive pruning ratio of 80%, SPARK maintains performance with less degradation than 5% compared to the based eviction method, demonstrating robustness and effectiveness. Our code will be available at \url{https://github.com/AMD-AIG-AIMA/AMD-Spark}.

Large language models (LLMs), such as ChatGPT, have achieved remarkable success across a wide range of fields. However, their trustworthiness remains a significant concern, as they are still susceptible to jailbreak attacks aimed at eliciting inappropriate or harmful responses. Most existing jailbreak attacks, nevertheless, mainly operate at the natural language level and rely on a single attack strategy, limiting their effectiveness in comprehensively assessing LLM robustness. In this paper, we propose Equacode, a novel multi-strategy jailbreak approach for large language models via equation-solving and code completion. This approach transforms malicious intent into a mathematical problem and then requires the LLM to solve it using code, leveraging the complexity of cross-domain tasks to divert the model's focus toward task completion rather than safety constraints. Experimental results show that Equacode achieves an average success rate of 91.19% on the GPT series and 97.62% across 5 state-of-the-art LLMs, all with only a single query. Further, ablation experiments demonstrate that EquaCode outperforms either the mathematical equation module or the code module alone. This suggests a strong synergistic effect, thereby demonstrating that multi-strategy approach yields results greater than the sum of its parts.

Although lip-to-speech synthesis (L2S) has achieved significant progress in recent years, current state-of-the-art methods typically rely on intermediate representations such as mel-spectrograms or discrete self-supervised learning (SSL) tokens. The potential of latent diffusion models (LDMs) in this task remains largely unexplored. In this paper, we introduce SLD-L2S, a novel L2S framework built upon a hierarchical subspace latent diffusion model. Our method aims to directly map visual lip movements to the continuous latent space of a pre-trained neural audio codec, thereby avoiding the information loss inherent in traditional intermediate representations. The core of our method is a hierarchical architecture that processes visual representations through multiple parallel subspaces, initiated by a subspace decomposition module. To efficiently enhance interactions within and between these subspaces, we design the diffusion convolution block (DiCB) as our network backbone. Furthermore, we employ a reparameterized flow matching technique to directly generate the target latent vectors. This enables a principled inclusion of speech language model (SLM) and semantic losses during training, moving beyond conventional flow matching objectives and improving synthesized speech quality. Our experiments show that SLD-L2S achieves state-of-the-art generation quality on multiple benchmark datasets, surpassing existing methods in both objective and subjective evaluations.

Pruning is a promising approach to reduce the high inference cost of large language models (LLMs), but it often comes at the expense of performance. Motivated by the "functional localization" theory in neuroscience, we hypothesize that LLMs contain task-specific expert activation paths, where specific subsets of neurons are co-activated for particular tasks. This structure allows selective activation to preserve task performance while improving inference efficiency. We introduce Sparse Expert Activation Pruning (SEAP), a training-free pruning method for large language models. SEAP identifies task-relevant activation paths by analyzing the clustering patterns of hidden states and neuron activations on a multi-task calibration dataset. Cross-task transfer evaluations confirm the existence of such expert activation structures. SEAP constructs task-aware pruning masks by leveraging a task-expert calibration dataset, which provides representative samples across diverse tasks to reveal their activation signatures. It then employs a lightweight task router to dynamically select relevant computation paths based on the input task. This design significantly reduces inference cost without compromising accuracy. Experimental results show that SEAP retains model performance with only a 1.5% drop on most tasks at 20% sparsity, and at 50% sparsity, it surpasses strong pruning baselines such as WandA and FLAP by over 20%. These results highlight SEAP as a scalable and effective solution for efficient LLM inference.

Retrieval-augmented generation (RAG) is widely adopted for knowledge-intensive tasks, but unverified external knowledge can pose risks such as data injection and retrieval pollution, leading to unexpected generation. Existing defenses rely on patch-based fixes, which limit generalization and increase system latency. To address these issues, we propose RAG2RAG, a framework-level security solution specifically designed for RAG. Inspired by human intuition to reason about what can and cannot be said during RAG phase, RAG2RAG augments the main RAG module with a lightweight RAG-based security expert module composed of two components: (1) a Detective that dynamically retrieves supporting evidence, and (2) a Judge that makes final decisions based on retrieved context. The main and expert modules operate in parallel without causing noticeable delays. Experiments across two languages, six domains, and seven types of poisoning attacks demonstrate that RAG2RAG overall achieves higher accuracy and lower attack success rates than seven mainstream baselines. Furthermore, it integrates seamlessly with various RAG architectures, offering efficient protection across diverse threat scenarios.

Parameter-Efficient finetuning (PEFT) enhances model performance on downstream tasks by updating a minimal subset of parameters. Representation finetuning (ReFT) methods further improve efficiency by freezing model weights and optimizing internal representations with fewer parameters than PEFT, outperforming PEFT on several tasks. However, ReFT exhibits a significant performance decline on mathematical reasoning tasks. To address this problem, the paper demonstrates that ReFT's poor performance on mathematical tasks primarily stems from its struggle to generate effective reasoning prefixes during the early inference phase. Moreover, ReFT disturbs the numerical encoding and the error accumulats during the CoT stage. Based on these observations, this paper proposes Bias-REstrained Prefix Representation FineTuning (BREP ReFT), which enhances ReFT's mathematical reasoning capability by truncating training data to optimize the generation of initial reasoning prefixes, intervening on the early inference stage to prevent error accumulation, and constraining the intervention vectors' magnitude to avoid disturbing numerical encoding. Extensive experiments across diverse model architectures demonstrate BREP's superior effectiveness, efficiency, and robust generalization capability, outperforming both standard ReFT and weight-based PEFT methods on the task of mathematical reasoning.

Large Language Models (LLMs) have demonstrated remarkable performance across a wide range of natural language processing tasks. However, temporal reasoning, particularly under complex temporal constraints, remains a major challenge. To this end, existing approaches have explored symbolic methods, which encode temporal structure explicitly, and reflective mechanisms, which revise reasoning errors through multi-step inference. Nonetheless, symbolic approaches often underutilize the reasoning capabilities of LLMs, while reflective methods typically lack structured temporal representations, which can result in inconsistent or hallucinated reasoning. As a result, even when the correct temporal context is available, LLMs may still misinterpret or misapply time-related information, leading to incomplete or inaccurate answers. To address these limitations, in this work, we propose Neuro-Symbolic Temporal Reasoning (NeSTR), a novel framework that integrates structured symbolic representations with hybrid reflective reasoning to enhance the temporal sensitivity of LLM inference. NeSTR preserves explicit temporal relations through symbolic encoding, enforces logical consistency via verification, and corrects flawed inferences using abductive reflection. Extensive experiments on diverse temporal question answering benchmarks demonstrate that NeSTR achieves superior zero-shot performance and consistently improves temporal reasoning without any fine-tuning, showcasing the advantage of neuro-symbolic integration in enhancing temporal understanding in large language models.

While large language models (LLMs) demonstrate emerging reasoning capabilities, current inference-time expansion methods incur prohibitive computational costs through exhaustive sampling. Through analyzing decoding trajectories, we observe that most next-token predictions align well with the golden output, except for a few critical tokens that lead to deviations. Inspired by this phenomenon, we propose a novel Hint-Practice Reasoning (HPR) framework that operationalizes this insight through two synergistic components: 1) a hinter (powerful LLM) that provides probabilistic guidance at critical decision points, and 2) a practitioner (efficient smaller model) that executes major reasoning steps. The framework's core innovation lies in Distributional Inconsistency Reduction (DIR), a theoretically-grounded metric that dynamically identifies intervention points by quantifying the divergence between practitioner's reasoning trajectory and hinter's expected distribution in a tree-structured probabilistic space. Through iterative tree updates guided by DIR, HPR reweights promising reasoning paths while deprioritizing low-probability branches. Experiments across arithmetic and commonsense reasoning benchmarks demonstrate HPR's state-of-the-art efficiency-accuracy tradeoffs: it achieves comparable performance to self-consistency and MCTS baselines while decoding only 1/5 tokens, and outperforms existing methods by at most 5.1% absolute accuracy while maintaining similar or lower FLOPs.

Visual Question Answering (VQA) requires models to reason over multimodal information, combining visual and textual data. With the development of continual learning, significant progress has been made in retaining knowledge and adapting to new information in the VQA domain. However, current methods often struggle with balancing knowledge retention, adaptation, and robust feature representation. To address these challenges, we propose a novel framework with adaptive memory allocation and global noise filtering called MacVQA for visual question answering. MacVQA fuses visual and question information while filtering noise to ensure robust representations, and employs prototype-based memory allocation to optimize feature quality and memory usage. These designs enable MacVQA to balance knowledge acquisition, retention, and compositional generalization in continual VQA learning. Experiments on ten continual VQA tasks show that MacVQA outperforms existing baselines, achieving 43.38% average accuracy and 2.32% average forgetting on standard tasks, and 42.53% average accuracy and 3.60% average forgetting on novel composition tasks.