AAAI 2026 论文列表

Speculative decoding accelerates LLM inference by utilizing otherwise idle computational resources during memory-to-chip data transfer. Current speculative decoding methods typically assume a considerable amount of available computing power, then generate a complex and massive draft tree using a small autoregressive language model to improve overall prediction accuracy. However, methods like batching have been widely applied in mainstream model inference systems as a superior alternative to speculative decoding, as they compress the available idle computing power. Therefore, performing speculative decoding with low verification resources and low scheduling costs has become an important research problem. We believe that more capable models that allow for parallel generation on draft sequences are what we truly need. Recognizing the fundamental nature of draft models to only generate sequences of limited length, we propose SpecFormer, a novel architecture that integrates unidirectional and bidirectional attention mechanisms. SpecFormer combines the autoregressive model’s ability to extract information from the entire input sequence with the parallel generation benefits of non-autoregressive models. This design eliminates the reliance on large prefix trees and achieves consistent acceleration, even in large-batch scenarios. Through lossless speculative decoding experiments across models of various scales, we demonstrate that SpecFormer sets a new standard for scaling LLM inference with lower training demands and reduced computational costs.

Training data detection is critical for enforcing copyright and data licensing, as Large Language Models (LLM) are trained on massive text corpora scraped from the internet. We present SPECTRA, a watermarking approach that makes training data reliably detectable even when it comprises less than 0.001% of the training corpus. SPECTRA works by paraphrasing text using an LLM and assigning a score based on how likely each paraphrase is, according to a separate scoring model. A paraphrase is chosen so that its score closely matches that of the original text, to avoid introducing any distribution shifts. To test whether a suspect model has been trained on the watermarked data, we compare its token probabilities against those of the scoring model. We demonstrate that SPECTRA achieves a consistent p-value gap of over nine orders of magnitude when detecting data used for training versus data not used for training, which is greater than all baselines tested. SPECTRA equips data owners with a scalable, deploy‑before‑release watermark that survives even large‑scale LLM training.

Retrieval-Augmented Generation (RAG) systems using Multimodal Large Language Models (MLLMs) show great promise for complex document understanding, yet their development is critically hampered by inadequate evaluation. Current benchmarks often focus on specific part of document RAG system and use synthetic data with incomplete ground truth and evidence labels, therefore failing to reflect real-world bottlenecks and challenges. To overcome these limitations, we introduce Double-Bench: a new large-scale, multilingual, and multimodal evaluation system that is able to produce fine-grained assessment to each component within document RAG systems. It comprises 3,276 documents (72,880 pages) and 5,168 single- and multi-hop queries across 6 languages and 4 document types with streamlined dynamic update support for potential data contamination issues. Queries are grounded in exhaustively scanned evidence pages and verified by human experts to ensure maximum quality and completeness. Our comprehensive experiments across 9 state-of-the-art embedding models, 4 MLLMs and 4 end-to-end document RAG frameworks demonstrate the gap between text and visual embedding models is narrowing, highlighting the need in building stronger document retrieval models. Our findings also reveal the over-confidence dilemma within current document RAG frameworks that tend to provide answer even without evidence support. We hope our fully open-source Double-Bench provide a rigorous foundation for future research in advanced document RAG systems. We plan to retrieve timely corpus and release new benchmarks on an annual basis.

Aligning Large Language Models (LLMs) with human preferences is critical, yet traditional fine-tuning methods are computationally expensive and inflexible. While test-time alignment offers a promising alternative, existing approaches often rely on distorted trajectory-level signals or inefficient sampling, fundamentally capping performance and failing to preserve the generative diversity of the base model. This paper introduces LLMdoctor, a novel framework for efficient test-time alignment that operates via a patient-doctor paradigm. It integrates token-level reward acquisition with token-level flow-guided preference optimization (TFPO) to steer a large, frozen patient LLM with a smaller, specialized doctor model. Unlike conventional methods that rely on trajectory-level rewards, LLMdoctor first extracts fine-grained, token-level preference signals from the patient model's behavioral variations. These signals then guide the training of the doctor model via TFPO, which establishes flow consistency across all subtrajectories, enabling precise token-by-token alignment while inherently preserving generation diversity. Extensive experiments demonstrate that LLMdoctor significantly outperforms existing test-time alignment methods and even surpasses the performance of full fine-tuning approaches like DPO.

Document clustering plays an important role in text mining and information retrieval. Existing methods primarily focus on document-intrinsic features, overlooking dataset-level features and consequently failing to construct superior representations. We propose a Contrastive Gaussian Fusion Network (CGFN) that can construct superior representations beyond the original documents. Specifically, CGFN fuses the Gaussian distributions of neighbor-derived information and intrinsic textual features in the latent space. By incorporating contrastive learning into the fusion process, our proposed method is able to learn high-quality representations while simultaneously mitigating noise and minimizing information loss. Experiments on four real-world datasets demonstrate that CGFN outperforms state-of-the-art methods, achieving superior clustering by robustly capturing holistic distributions and neighbor patterns.

The impressive performance of large language models (LLMs) also brings inherent toxicity risks, prompting the need for effective detoxification to support responsible deployment. Prevailing methods generally follow an inflexible model-specific fashion, addressing only individual models or model families. Moreover, overlooking the underlying toxic risks involved in the input prefix can lead to toxic accumulation during autoregressive generation. Existing methods rely on external strong attribute interventions to address this issue, which further exacerbates contextual semantic inconsistencies and makes it difficult to balance toxicity efficacy and generation quality. To address these concerns, we propose a novel Model-Agnostic Adaptive Detoxification (MAAD) framework. To address accumulating toxicity, we present prefix heuristics that serve as contextual signals, guiding the base LLM toward safer generation. Along this line, we construct an antidote dataset to support a lightweight model, Detoxifier, which steers the base LLM to make in-scope and reliable detoxifying distribution adjustments while preserving fluency and contextual understanding. Designed as an easy-to-deploy module, Detoxifier requires a small amount of data and can be seamlessly applied to various base LLMs with one-off training. Since over-purifying often reduces diversity, we also propose a dynamic truncation method called CW-cutoff sampling to trade off language model quality and diversity. Extensive experiments demonstrate that MAAD strikes a better balance between detoxification effectiveness and generation quality, while also maintaining model utility.

The release of open-weight large language models (LLMs) creates a tension between advancing accessible research and preventing misuse, such as malicious fine-tuning to elicit harmful content. Current safety measures struggle to preserve the general capabilities of the LLM while resisting a determined adversary with full access to the model's weights and architecture, who can use full-parameter fine-tuning to erase existing safeguards. To address this, we introduce AntiDote, a bi-level optimization procedure for training LLMs to be resistant to such tampering. AntiDote involves an auxiliary adversary hypernetwork that learns to generate malicious Low-Rank Adaptation (LoRA) weights conditioned on the defender model's internal activations. The defender LLM is then trained with an objective to nullify the effect of these adversarial weight additions, forcing it to maintain its safety alignment. We validate this approach against a diverse suite of 52 red-teaming attacks, including jailbreak prompting, latent space manipulation, and direct weight-space attacks. AntiDote is upto 27.4% more robust against adversarial attacks compared to both tamper-resistance and unlearning baselines. Crucially, this robustness is achieved with a minimal trade-off in utility, incurring a performance degradation of upto less than 0.5% across capability benchmarks including MMLU, HellaSwag, and GSM8K. Our work offers a practical and compute efficient methodology for building open-weight models where safety is a more integral and resilient property.

To develop general-purpose collaborative agents, humans need reliable AI systems that can (1) adapt to new domains and (2) transparently reason with uncertainty to allow for verification and correction. Black-box models demonstrate powerful data processing abilities but do not satisfy these criteria due to their opaqueness, domain specificity, and lack of uncertainty awareness. We introduce Bonsai, a compositional and probabilistic reasoning system that generates adaptable inference trees by retrieving relevant grounding evidence and using it to compute likelihoods of sub-claims derived from broader natural language inferences. Bonsai's reasoning power is tunable at test-time via evidence scaling and it demonstrates reliable handling of varied domains including transcripts, photographs, videos, audio, and databases. Question-answering and human alignment experiments demonstrate that Bonsai matches the performance of domain-specific black-box methods while generating interpretable, grounded, and uncertainty-aware reasoning traces.

Adapting large language models (LLMs) to new languages is an expensive and opaque process. Understanding how language models acquire new languages and multilingual abilities is key to achieve efficient adaptation. Prior work on multilingual interpretability research focuses primarily on how trained models process multilingual instructions, leaving unexplored the mechanisms through which they acquire new languages during training. We investigate these training dynamics on decoder-only transformers through the lens of two functional cognitive specializations: language perception (input comprehension) and production (output generation). Through experiments on low-resource languages, we demonstrate how perceptual and productive specialization emerges in different regions of a language model by running layer ablation sweeps from the model’s input and output directions. Based on the observed specialization patterns, we propose CogSym, a layer-wise heuristic that enables effective adaptation by exclusively finetuning a few early and late layers. We show that tuning only the 25% outermost layers achieves downstream task performance within 2–3% deviation from the full finetuning baseline. Unlike similar layer-selection methods, the proposed method requires no extra data or computation while retaining comparable performance, which is especially beneficial for low-resource languages. CogSym yields consistent performance with adapter methods such as LoRA, showcasing generalization beyond full finetuning. These findings provide insights to better understand how LLMs learn new languages and push toward accessible and inclusive language modeling.

Computational humor is a frontier for creating advanced and engaging natural language processing (NLP) applications, such as sophisticated dialogue systems. While previous studies have benchmarked the humor capabilities of Large Language Models (LLMs), they have often relied on single-dimensional evaluations, such as judging whether something is simply ``funny.'' This paper argues that a multifaceted understanding of humor is necessary and addresses this gap by systematically evaluating LLMs through the lens of Oogiri, a form of Japanese improvisational comedy games. To achieve this, we expanded upon existing Oogiri datasets with data from new sources and then augmented the collection with Oogiri responses generated by LLMs. We then manually annotated this expanded collection with 5-point absolute ratings across six dimensions: Novelty, Clarity, Relevance, Intelligence, Empathy, and Overall Funniness. Using this dataset, we assessed the capabilities of state-of-the-art LLMs on two core tasks: their ability to generate creative Oogiri responses and their ability to evaluate the funniness of responses using a six-dimensional evaluation. Our results show that while LLMs can generate responses at a level between low- and mid-tier human performance, they exhibit a notable lack of Empathy. This deficit in Empathy helps explain their failure to replicate human humor assessment. Correlation analyses of human and model evaluation data further reveal a fundamental divergence in evaluation criteria: LLMs prioritize Novelty, whereas humans prioritize Empathy. We release our annotated corpus to the community to pave the way for the development of more emotionally intelligent and sophisticated conversational agents.

ColBERT introduced a late interaction mechanism that independently encodes queries and documents using BERT, and computes similarity via fine-grained interactions over token-level vector representations. This design enables expressive matching while allowing efficient computation of scores, as the multi-vector document representations could be pre-computed offline. ColBERT models distance using a Chamfer-style function: for each query token, it selects the closest document token and sums these distances across all query tokens. In our work, we explore enhancements to the Chamfer distance function by computing a weighted sum over query token contributions, where weights reflect the token importance. Empirically, we show that this simple extension, requiring only token-weight training while keeping the multi-vector representations fixed, further enhances the expressiveness of late interaction multi-vector mechanism. In particular, on the BEIR benchmark, our method achieves an average improvement of 1.28% in Recall@10 in the zero-shot setting using IDF-based weights, and 3.66% through few-shot fine-tuning.

Large Language Models (LLMs) have achieved remarkable success in instruction-following and dialogue tasks, yet aligning them with human preferences remains a critical challenge. Recent advances such as Direct Preference Optimization (DPO) simplify the alignment pipeline by bypassing explicit reward modeling, but they often suffer from suboptimal reward margin distributions, leading to weak supervision signals and reduced discriminative capacity. In this work, we propose Reward Margin Optimization (RMO), a framework that reshapes reward margin distributions during training to improve alignment performance. RMO comprises three components: (1) a Dual Denoising Filtering strategy that filters ambiguous and noisy preference pairs based on reward margin dynamics; (2) Batch Margin Diversification, which maximizes intra-batch margin variance to enhance learning signal diversity; and (3) Pairwise Margin Amplification, an auxiliary regularization term that encourages larger margins between preferred and dispreferred responses. Extensive experiments on multiple LLMs and datasets demonstrate that RMO consistently improves win rates over strong baselines such as DPO and SimPO, while remaining compatible with various preference-based optimization methods. Our results highlight the critical role of reward margin distribution in preference alignment and establish RMO as an effective and scalable enhancement to existing alignment techniques.

Standard single-turn, static benchmarks fall short in evaluating the nuanced capabilities of Large Language Models (LLMs) on complex tasks such as software engineering. In this work, we propose a novel interactive evaluation framework that assesses LLMs on multi-requirement programming tasks through structured, feedback-driven dialogue. Each task is modeled as a requirement dependency graph, and an "interviewer" LLM, aware of the ground-truth solution, provides minimal, targeted hints to an "interviewee" model to help correct errors and fulfill target constraints. This dynamic protocol enables fine-grained diagnostic insights into model behavior, uncovering strengths and systematic weaknesses that static benchmarks fail to measure. We build on DevAI, a benchmark of 55 curated programming tasks, by adding ground-truth solutions and evaluating the relevance and utility of interviewer hints through expert annotation. Our results highlight the importance of dynamic evaluation in advancing the development of collaborative code-generating agents.

Automatic speech recognition (ASR) systems have achieved remarkable performance in common conditions but often struggle to leverage long-context information in contextualized scenarios that require domain-specific knowledge, such as conference presentations. This challenge arises primarily due to constrained model context windows and the sparsity of relevant information within extensive contextual noise. To solve this, we propose the SAP^2 method, a novel framework that dynamically prunes and integrates relevant contextual keywords in two stages. Specifically, each stage leverages our proposed Speech-Driven Attention-based Pooling mechanism, enabling efficient compression of context embeddings while preserving speech-salient information. Experimental results demonstrate state-of-the-art performance of SAP^2 on the SlideSpeech and LibriSpeech datasets, achieving word error rates (WER) of 7.71% and 1.12%, respectively. On SlideSpeech, our method notably reduces biased keyword error rates (B-WER) by 41.1% compared to non-contextual baselines. SAP^2 also exhibits robust scalability, consistently maintaining performance under extensive contextual input conditions on both datasets.

Generative models have shown remarkable performance in speech enhancement (SE), achieving superior perceptual quality over traditional discriminative approaches. However, existing generative SE approaches often overlook the risk of hallucination under severe noise, leading to incorrect spoken content or inconsistent speaker characteristics, which we term linguistic and acoustic hallucinations, respectively. We argue that linguistic hallucination stems from models' failure to constrain valid phonological structures and it is a more fundamental challenge. While language models (LMs) are well-suited for capturing the underlying speech structure through modeling the distribution of discrete tokens, existing approaches are limited in learning from noise-corrupted representations, which can lead to contaminated priors and hallucinations. To overcome these limitations, we propose the Phonologically Anchored Speech Enhancer (PASE), a generative SE framework that leverages the robust phonological prior embedded in the pre-trained WavLM model to mitigate hallucinations. First, we adapt WavLM into a denoising expert via representation distillation to clean its final-layer features. Guided by the model's intrinsic phonological prior, this process enables robust denoising while minimizing linguistic hallucinations. To further reduce acoustic hallucinations, we train the vocoder with a dual-stream representation: the high-level phonetic representation provides clean linguistic content, while a low-level acoustic representation retains speaker identity and prosody. Experimental results demonstrate that PASE not only surpasses state-of-the-art discriminative models in perceptual quality, but also significantly outperforms prior generative models with substantially lower linguistic and acoustic hallucinations.

With the increasing commercialization of the latent diffusion-based text-to-audio generation, model attribution has become a critical challenge. Embedding watermarks in generated audio is an effective way to distinguish synthetic from natural audio. However, existing watermarking methods often suffer from limited robustness or require additional training, limiting their scalability in practical applications. In this paper, we propose an anchor-based inversion optimization framework. The method embeds a watermark into the model's initial latent vector, designated as a pivotal anchor, and extracts the watermark through inversion. To mitigate error accumulation and enhance robustness during inversion, we leverage the temporal consistency and distributional similarity of diffusion models, formulating watermark extraction as a time-series optimization problem. Specifically, given a suspicious audio sample and a candidate model with a predefined anchor, we first perform unguided denoising diffusion on the anchor to generate an intermediate latent trajectory as the anchor sequence. Then, we optimize the inversion process to align the inverted trajectory with the anchor sequence, thereby reducing accumulated errors. During optimization, we adopt Soft Dynamic Time Warping as the loss function. Its flexible temporal alignment capability ensures that correct attribution is achieved only when the anchor matches the target audio. Experimental results show that our method enables training-free attribution while preserving audio quality and achieving strong robustness.

Process or step-wise supervision has played a crucial role in advancing complex multi-step reasoning capabilities of Large Language Models (LLMs). However, efficient, high-quality automated process annotation remains a significant challenge. To address this, we introduce Single-Pass Annotation with Reference-Guided Evaluation (SPARE), a novel structured framework that enables efficient per-step annotation by jointly aligning solution steps to reference solutions and determine its accuracy with explicit reasoning in single generation. We demonstrate SPARE's effectiveness across four diverse datasets spanning mathematical reasoning (GSM8K, MATH), multi-hop question answering (MuSiQue-Ans), and spatial reasoning (SpaRP), showing consistent improvements in two applications: (1) training Process Reward Models (PRMs) for ranking and aggregating multiple generations, and (2) fine-tuning models via offline reinforcement learning for greedy decoding. On PROCESSBENCH, SPARE demonstrates data-efficient out-of-distribution generalization, using only ~16% of training samples compared to human-labeled and other synthetically trained baselines. Additionally, it achieves competitive performance with MCTS-based methods while offering 2.3x speedup in terms of total token count. Manual analysis reveals complementary precision-recall characteristics with MCTS approaches, suggesting potential for ensemble methods. These results establish SPARE as a practical and scalable solution for automatic process supervision in LLM reasoning.

Human-interaction-involved applications underscore the need for Multi-modal Sentiment Analysis (MSA). Although many approaches have been proposed to address the subtle emotions in different modalities, the power of explanations and temporal alignments is still underexplored. Thus, this paper proposes the Text-routed sparse mixture-of-Experts model with eXplanation and Temporal alignment for MSA (TEXT). TEXT first augments explanations for MSA via Multi-modal Large Language Models (MLLM), and then novelly aligns the representations of audio and video through a temporality-oriented neural network block. TEXT aligns different modalities with explanations and facilitates a new text-routed sparse mixture-of-experts with gate fusion. Our temporal alignment block merges the benefits of Mamba and temporal cross-attention. As a result, TEXT achieves the best performance across four datasets among all tested models, including three recently proposed approaches and three MLLMs. TEXT wins on at least four metrics out of all six metrics. For example, TEXT decreases the mean absolute error to 0.353 on the CH-SIMS dataset, which signifies a 13.5% decrement compared with recently proposed approaches.

Large Language Models (LLMs) surprised the world with their ability to mimic humans in writing and are starting to be used as simulations of human writers for various kinds of linguistic analyses. However, these analyses rest on the belief that LLMs are good density models that accurately capture the underlying probability distribution of the language. In this paper, we question this basic assumption and try to evaluate language models on their density modelling capabilities. Since a ground truth does not exist for the probability distribution of any natural language, we come up with a synthetic language made up of decimal numbers written in words in English. We train language models from scratch on various probability distributions over this synthetic language and compare the distributions learned by the models with the original distributions. Experiments show that language models can learn underlying probability distributions across a wide range of cases, but they fail when those distributions depend on deep semantic properties of numbers that cannot be inferred from syntactic patterns. Additionally, we observed a strong bias in the models towards numbers that frequently occur as substrings within other numbers. This suggests that such a bias possibly exists in real-world natural language models as well, and negatively impacts downstream tasks and analyses that rely on model-generated probabilities.