论文检索

Deep Learning · Large Language Models

Haoyu Wang, yifan shang, Zhongxiang Sun, Weijie Yu, Xiao Zhang, Jun Xu

Continual Pre-Training (CPT) is essential for enabling Language Models (LMs) to integrate new factual knowledge without erasing old. While classical CPT techniques like data replay have become the standard paradigm, the mechanisms underlying how LMs acquire and retain facts over time, termed as continual Factual Knowledge Acquisition (cFKA), remain unclear. In this work, we present a theoretical framework that characterizes the training dynamics of cFKA using a single-layer Transformer with linear attention, offering a unified explanation for the behavior of popular CPT methods. Our analysis reveals that regularization-based methods merely adjust the convergence rate of parameters without altering the inherent forgetting tendency, whereas data replay methods shift convergence dynamics and stabilize pretrained knowledge. Building on these insights, we propose a novel generative data replay approach, called Selecting Tokens via attentiOn Contribution (STOC), which identifies influential factual snippets to guide replay generation. Extensive experiments on both synthetic and real-world datasets validate our theoretical findings and demonstrate that STOC effectively enhances cFKA by mitigating catastrophic forgetting.

Theory · Domain Adaptation and Transfer Learning

Zhexiao Lin, Yuanyuan Li, Neeraj Sarna, Yuanyuan Gao, Michael Berger

Large language models have achieved impressive performance across diverse tasks. However, their tendency to produce overconfident and factually incorrect outputs, known as hallucinations, poses risks in real world applications. Conformal prediction provides finite-sample, distribution-free coverage guarantees, but standard conformal prediction breaks down under domain shift, often leading to under-coverage and unreliable prediction sets. We propose a new framework called Domain-Shift-Aware Conformal Prediction (DS-CP). Our framework adapts conformal prediction to large language models under domain shift, by systematically reweighting calibration samples based on their proximity to the test prompt, thereby preserving validity while enhancing adaptivity. Our theoretical analysis and experiments on the MMLU benchmark demonstrate that the proposed method delivers more reliable coverage than standard conformal prediction, especially under substantial distribution shifts, while maintaining efficiency. This provides a practical step toward trustworthy uncertainty quantification for large language models in real-world deployment.

Reinforcement Learning · Planning

Shiyu Li, Yifan Wang, Peiming Li, Zheng Wei, Yang Tang

Search agents powered by Large Language Models have demonstrated significant potential in tackling knowledge-intensive tasks. Reinforcement learning has emerged as a powerful paradigm for training these agents to perform complex, multi-step reasoning. However, prior RL-based methods often rely on sparse or rule-based rewards, which can lead agents to commit to suboptimal or erroneous reasoning paths without the ability to recover. To address these limitations, we propose **ReSeek**, a self-correcting framework enabling search agents to recover from erroneous search paths during an episode. By invoking a special **JUDGE** action, the agent can judge the information and re-plan its search strategy. To guide this process, we design a dense, instructive process reward function, which decomposes into a correctness reward for retrieving factual information and a utility reward for finding information genuinely useful for the query. Additionally, to mitigate the risk of data contamination in existing datasets, we introduce **FictionalHot**, a contamination-free benchmark requiring complex reasoning. Experiments show ReSeek significantly outperforms SOTA baselines in task success and path faithfulness.

Deep Learning · Large Language Models

Zhuofan Shi, Ming Ma, ZekunYao, Fangkai Yang, Jue Zhang, Dongge Han, Victor Ruehle, Qingwei Lin, Saravan Rajmohan, Dongmei Zhang

Open-Ended Deep Research (OEDR) pushes LLM agents beyond short-form QA toward long-horizon workflows that iteratively search, connect, and synthesize evidence into structured reports. However, existing OEDR agents largely follow either linear "search-then-generate" accumulation or outline-centric planning. The former suffers from lost-in-the-middle failures as evidence grows, while the latter relies on the LLM to implicitly infer knowledge gaps from the outline alone, providing weak supervision for identifying missing relations and triggering targeted exploration. We present DualGraph memory, an architecture that separates what the agent knows from how it writes. DualGraph maintains two co-evolving graphs: an Outline Graph (OG), and a Knowledge Graph (KG), a semantic memory that stores fine-grained knowledge units, including core entities, concepts, and their relations. By analyzing the KG topology together with structural signals from the OG, DualGraph generates targeted search queries, enabling more efficient and comprehensive iterative knowledge-driven exploration and refinement. Across DeepResearch Bench, DeepResearchGym, and DeepConsult, DualGraph consistently outperforms state-of-the-art baselines in report depth, breadth, and factual grounding; for example, it reaches a 53.08 RACE score on DeepResearch Bench with GPT-5. Moreover, ablation studies confirm the central role of the dual-graph design. DualGraph code is available at https://anonymous.4open.science/r/DualGraph-2536.

Deep Learning · Large Language Models

Jiayi Zhang, Simon Yu, Derek Chong, Anthony Sicilia, Michael Tomz, Christopher Manning, Weiyan Shi

Post-training alignment often reduces LLM diversity, leading to a phenomenon known as mode collapse. Unlike prior work that attributes this effect to algorithmic limitations, we identify a fundamental, pervasive data-level driver: typicality bias in preference data, whereby annotators systematically favor familiar text as a result of well-established findings in cognitive psychology. We formalize this bias theoretically, verify it empirically on preference datasets, and show that it plays a central role in mode collapse. Motivated by this analysis, we introduce Verbalized Sampling (VS), a simple, training-free prompting strategy to circumvent mode collapse. VS prompts the model to verbalize a probability distribution over a set of responses (e.g., "Generate 5 jokes about coffee and their corresponding probabilities"), which relieves the pressure to produce a single "typical" answer. Experiments show that VS significantly improves performance across creative writing (poems, stories, jokes), social dialogue simulation, synthetic data generation, and open-ended QA, without sacrificing safety and factual accuracy. For instance, in creative writing, VS increases diversity by 1.6-2.1x compared to direct prompting. We further observe an emergent trend that more capable models benefit more from VS. In sum, our work provides a new data-centric perspective on mode collapse and a practical inference-time remedy that helps unlock pre-trained generative diversity.

Social Aspects · Alignment

Itai Shapira, Gerdus Benade, Ariel Procaccia

Large language models often exhibit increased sycophantic behavior after preference-based post-training, showing a stronger tendency to affirm a user’s stated or implied belief even when this conflicts with factual accuracy or sound judgment. We present a formal analysis of how alignment from human feedback can increase this failure mode by identifying an explicit amplification mechanism that causally links optimization against a learned reward to bias in the human preference data used for alignment. We show that the direction of behavioral drift is determined by a covariance under the base policy between endorsing the belief signal in the prompt and the learned reward, and that the first-order effect reduces to a simple mean-gap condition. We then analyze reward learning from pairwise comparisons under random utility models like Bradley–Terry and characterize when bias in human annotators’ preferences induces this reward gap. Next, we propose a training-time intervention designed to neutralize the amplification mechanism itself. Among all post-trained policies that prevent sycophantic behavior from increasing, we characterize the unique policy closest in KL divergence to the unconstrained post-trained policy, and derive the corresponding minimal reward correction as a closed-form agreement penalty. Computational experiments find that reward gaps are common and cause behavioral drift in all the configurations considered.

Applications · Language, Speech and Dialog

Wenxuan Huang, Yu Zeng, Qiuchen Wang, Zhen Fang, Shaosheng Cao, Zheng Chu, Qingyu Yin, Shuang Chen, Zhenfei Yin, Lin Chen 等

Multimodal large language models (MLLMs) have achieved remarkable success across a broad range of vision tasks. However, constrained by the capacity of their internal world knowledge, prior work has proposed augmenting MLLMs by ``reasoning-then-tool-call'' for visual and textual search engines to obtain substantial gains on tasks requiring extensive factual information. However, these approaches typically define multimodal search in a naive setting, assuming that a single full-level or entity-level image query and few text query suffices to retrieve the key evidence needed to answer the question, which is unrealistic in real-world scenarios with substantial visual noise. Moreover, they are often limited in the reasoning depth and search breadth, making it difficult to solve complex questions that require aggregating evidence from diverse visual and textual sources. Building on this, we propose Vision-DeepResearch, which proposes one new multimodal deep-research paradigm, i.e., performs multi-turn, multi-entity and multi-scale visual and textual search to robustly hit real-world search engines under heavy noise. Our Vision-DeepResearch supports dozens of reasoning steps and hundreds of engine interactions, while internalizing deep-research capabilities into the MLLM via cold-start supervision and RL training, resulting in a strong end-to-end multimodal deep-research MLLM. It substantially outperforming existing multimodal deep-research MLLMs, and workflows built on strong closed-source foundation model such as GPT-5, Gemini-2.5-pro and Claude-4-Sonnet.

Deep Learning · Large Language Models

Dhrupad Bhardwaj, Julia Kempe, Tim G. J. Rudner

To deploy large language models (LLMs) in high-stakes application domains that require substantively accurate responses to open-ended prompts, we need reliable, computationally inexpensive methods that assess the trustworthiness of long-form responses generated by LLMs. However, existing approaches often rely on claim-by-claim fact-checking, which is computationally expensive and brittle in long-form responses to open-ended prompts. In this work, we introduce semantic isotropy—the degree of uniformity across normalized text embeddings on the unit sphere—and use it to assess the trustworthiness of long-form responses generated by LLMs. To do so, we generate several long-form responses, embed them, and estimate the level of semantic isotropy of these responses as the angular dispersion of the embeddings on the unit sphere. We find that higher semantic isotropy—that is, greater embedding dispersion—reliably signals lower factual consistency across samples. Our approach requires no labeled data, no fine-tuning, and no hyperparameter selection, and can be used with open- or closed-weight embedding models. Across multiple domains, our method consistently outperforms existing approaches in predicting nonfactuality in long-form responses using only a handful of samples—offering a practical, low-cost approach for integrating trust assessment into real-world LLM workflows.

Deep Learning · Foundation Models

Xinnan Dai, Kai Yang, cheng Luo, Shenglai Zeng, Kai Guo, Jiliang Tang

Reasoning hallucinations in large language models (LLMs) often appear as fluent yet unsupported conclusions that violate either the given context or underlying factual knowledge. Although such failures are widely observed, the mechanisms by which decoder-only Transformers produce them remain poorly understood. We model next-token prediction as a graph search process over an underlying graph, where entities correspond to nodes and learned transitions form edges. From this perspective, contextual reasoning is a constrained search over a sampled subgraph (intrinsic reasoning), while context-free queries rely on memorized structures in the underlying graph (extrinsic reasoning). We show that reasoning hallucinations arise from two fundamental mechanisms: path reuse, where memorized knowledge overrides contextual constraints during early training, and path compression, where frequently traversed multi-step paths collapse into shortcut edges in later training. Together, these mechanisms provide a unified explanation for reasoning hallucinations in LLMs and connected to well-known behaviors observed in downstream applications.

Deep Learning · Large Language Models

Nathan Hittesdorf, Marco Salzetta, Lu Cheng

Large Language Models (LLMs) frequently hallucinate, limiting their reliability in critical applications. Conformal Prediction (CP) addresses this by calibrating error rates on held-out data to provide statistically valid confidence guarantees. Recent work extends CP to LLM factuality: outputs are decomposed into subclaims, each assigned a risk score, and a calibrated threshold filters out risky claims to guarantee hallucination rates below a user-specified level (e.g., 10%). While prior methods treat claims independently, Coherent Factuality extends to multi-step reasoning by representing outputs as dependency graphs and jointly validating claims with their logical ancestors. A key limitation is that Coherent Factuality is not differentiable, requiring hand-crafted scorers that at high reliability levels remove nearly 60% of true claims. We introduce Differentiable Coherent Factuality (DCF), a fully differentiable relaxation that enables learning improved scorers while provably recovering the original algorithm's guarantees. Experiments on two reasoning datasets demonstrate DCF achieves up to 141% improvement in claim retention while maintaining reliability guarantees, representing a significant step towards reliable conformal LLM systems.

Yan Zhang, Zhijin Qin, Feng Xu, guiguang ding, Jungong Han

Despite rapid progress in Grounded Visual Reasoning (GVR) with MLLMs and RL-style fine-tuning, existing approaches often lack effective learning signals for intermediate grounding decisions and are prone to shortcut solutions. In this work, we explicitly decompose GVR into Evidence Generation followed by Counterfactual Answer Reasoning, and formalize this structure as a Causal Grounding Graph (CGG) in which the generated evidence acts as a causal mediator. Building on this formulation, we propose CofactGVR, which estimates the mediator’s utility via a matched counterfactual intervention that perturbs the predicted region while keeping the original image–question context fixed. The factual–counterfactual reward gap yields a principled intermediate bonus, selectively assigned to high-quality factual rollouts to promote evidence-faithful reasoning. To further stabilize and efficiently exploit this causal training signal, we incorporate a Quantile-filtered Prioritized Advantage Sampling (Q-PAS) strategy that preferentially updates on trajectories with high-magnitude advantages while filtering low-signal samples. Extensive experiments across GVR benchmarks show consistent improvements, indicating that CofactGVR strengthens reliance on informative visual evidence under controlled interventions. The source codes will be publicly available.

Deep Learning · Attention Mechanisms

Wei-Yao Wang, Zhao Wang, Helen Suzuki, Yoshiyuki Kobayashi

Recent Multimodal Large Language Models (MLLMs) have demonstrated significant progress in perceiving and reasoning over multimodal inquiries, ushering in a new research era for foundation models. However, vision-language misalignment in MLLMs has emerged as a critical challenge, where the textual responses generated by these models are not factually aligned with the given text-image inputs. Existing efforts to address vision-language misalignment have focused on developing specialized vision-language connectors or leveraging visual instruction tuning from diverse domains. In this paper, we tackle this issue from a fundamental yet unexplored perspective by revisiting the core architecture of MLLMs. Most MLLMs are typically built on decoder-only LLMs consisting of a causal attention mechanism, which *limits the ability of the earlier modalities (e.g., images) to incorporate information from the latter modalities (e.g., text)*. To address this problem a MLLM that unlocks causal attention into our proposed modality-mutual attention (MMA) to enable image tokens to attend to text tokens. This simple yet effective design allows MMA to achieve state-of-the-art performance in 12 multimodal understanding benchmarks (**+6.2\% on average across 3 LLMs backbones**) without introducing additional parameters. Our MMA design is intended to be generic, allowing for applications across various modalities, and scalable to accommodate diverse multimodal scenarios.

Social Aspects · Alignment

Sougata Saha, Madhur Jindal, Saurabh Kumar Pandey, Mahardika Ihsani, Alham Fikri Aji, Monojit Choudhury

Most existing cultural evaluation frameworks for large language models (LLMs) focus on matching model outputs to ground-truth answers, primarily measuring factual cultural awareness. This overlooks whether models internalize broader cultural structure and pluralism. We introduce a spectral-analysis-based framework that captures large-scale macrostructural patterns in models' cultural knowledge and evaluate eight LLMs across nine cultural domains spanning all five of Newmark's cultural dimensions and 170 countries. Comparing with human data, we find that instruction-tuned models align more closely with human cultural structure than older models, while increased model size does not consistently improve performance. Finally, simulation-based experiments show that our proposed spectral metric better predicts a model's ability to serve users from unfamiliar cultural backgrounds than existing ones.

Deep Learning · Large Language Models

Wuyang Zhang, Shichao Pei

Retrieval-Augmented Generation (RAG) improves factual grounding in large language models but suffers from substantial latency due to synchronous retrieval. While recent work explores asynchronous retrieval, existing approaches rely on heuristic coordination between retrieval and generation and assume stable information demands during decoding that often break in complex, multi-domain settings. In this paper, we propose an advanced asynchronous retrieval framework that enables predictive prefetching aligned with evolving information needs. The framework explicitly predicts when retrieval should be triggered and what information should be retrieved using three components, a retrieval predictor, a context monitor, and a query generator, by exploiting semantic precursors in generation dynamics that emerge several tokens before uncertainty becomes critical. Experiments on multiple benchmarks demonstrate up to 43.5\% end-to-end latency reduction and 62.4\% improvement in time-to-first-token, while maintaining answer quality comparable to synchronous RAG baselines.

General Machine Learning · Evaluation

Ilija Subasic, Andrew Rabinovich, Zhao Chen

As Large Language Models (LLMs) are increasingly deployed to serve open-ended, multi-turn interactions, evaluating conversational quality at human scale has become a central challenge. Existing evaluation frameworks built for summarization, translation, or short-form QA tasks fall short of adequately measuring the consistency of human-scale dialogue, especially when derivation and validation of these metrics themselves often rely on synthetic rather than human sources. We fill the gap by introducing UPHELD (Utility & Planning Human-Scale Evaluated Long Dialogues), a large, reference-full benchmark for evaluating human-scale conversational ability beyond factual correctness. UPHELD consists of hundreds of complete human-to-human dialogues authored by professional script writers, with realistic turn densities and 36,000+ per-turn human annotations across 10,000+ expert-generated dialogue turns. Using UPHELD, we systematically evaluate classical automatic metrics and reference-free LLM-as-a-judge approaches, and find them unreliable when correlated with expert human judgment. Building off this analysis, we use UPHELD to develop a Mixture-of-Judges framework that combines multiple evaluative signals and improves correlation with human assessments by approximately 30%. Overall, UPHELD provides a robust, human-grounded foundation for evaluating long, human-scale conversational intelligence that fills a crucial gap in the pre-existing LLM dataset landscape

Reinforcement Learning · Policy Search

Abdulhady abas, Fatemeh Daneshfar, Seyedali Mirjalili, Mourad Oussalah

Aligning large language models (LLMs) with human preferences is commonly done via reinforcement learning from human feedback (RLHF) with Proximal Policy Optimization (PPO) or, more simply, via Direct Preference Optimization (DPO). While DPO is stable and RL-free, it treats preferences as flat winner vs. loser signals and is sensitive to noisy or brittle preferences arising from fragile chains of thought. We propose TUR-DPO, a topology- and uncertainty-aware variant of DPO that rewards how answers are derived, not only what they say, by eliciting lightweight reasoning topologies and combining semantic faithfulness, utility, and topology quality into a calibrated uncertainty signal. A small learnable reward is factorized over these signals and incorporated into an uncertainty-weighted DPO objective that remains RL-free and relies only on a fixed or moving reference policy. Empirically, across open 7–8B models and benchmarks spanning mathematical reasoning, factual question answering, summarization, and helpful/harmless dialogue, TUR-DPO improves judge win-rates, faithfulness, and calibration relative to DPO while preserving training simplicity and avoiding online rollouts. We further observe consistent gains in multimodal and long-context settings, and show that TUR-DPO matches or exceeds PPO on reasoning-centric tasks while maintaining operational simplicity.

Social Aspects · Accountability, Transparency, and Interpretability

Jacqueline He, Jonathan Hayase, Scott Yih, Sewoong Oh, Luke Zettlemoyer, Pang Wei Koh

Modern language models (LMs) tend to memorize portions of their training data and reproduce verbatim spans. When the underlying sources are sensitive or copyright-protected, such reproduction raises issues of consent and compensation for creators and compliance risks for developers. We propose Proximal Decoding, a plug-and-play inference-time method for suppressing verbatim reproduction: it enables decoding from any risky LM trained on mixed-license data by keeping generation in bounded proximity to a permissively trained safe LM. Proximal Decoding does so by adaptively allocating a user-chosen information budget over the generation trajectory and enforcing per-step constraints that yield a sequence-level guarantee, enabling a tunable risk–utility trade-off. To make Proximal Decoding practically useful, we introduce a new permissively trained safe model (Comma 1.7B), as well as Proximal$\_{\mathrm{Byte}}$, a byte-level variant of our method that enables cross-vocabulary fusion via the ByteSampler (Hayase et al., 2025) framework. We evaluate our methods across six model pairs on long-form evaluations of copyright risk and utility. Proximal and Proximal$\_{\mathrm{Byte}}$ define a new Pareto frontier, preserving near-original fluency and factuality while eliminating up to 75\% of the measurable copying gap (averaged over six copying metrics) between the risky baseline and a safe reference, at a modest inference overhead.

Deep Learning · Large Language Models

Zheng Wang, Kaixuan Zhang, Wanfang Chen, Jingwen Zhang, Xiaonan Lu

Sequential editing of structured knowledge in large language models allows targeted factual updates without retraining, yet existing methods often rely on complex regularization or constraint mechanisms whose necessity remains unclear. In this work, we systematically investigate the mechanisms underlying effective and stable sequential editing. Specifically, we first analyze the empirical success of AlphaEdit and establish, via a rigorous optimization analysis, the formal equivalence between one-time and sequential editing. Building on this insight, we generalize the equivalence to a broader class of editing objectives, demonstrating that stability emerges naturally from properly accounting for accumulated editing constraints, rather than from specialized regularization or null-space operations. We empirically confirm that many commonly used regularization strategies are unnecessary for reliable sequential updates. Furthermore, we extend our framework to handle conflicting edits, ensuring robust and consistent behavior under contradictory updates. Ultimately, our work provides Ariadne’s thread through the labyrinth of sequential editing, charting a path toward simpler, more interpretable, and dependable knowledge updates.

Siqi Kou, Jiachun Jin, Jiayin Chen, Ye Ma, Yugang Wang, Quan Chen, Peng Jiang, Xiao Yang, Jun Zhu, Kai Yu 等

Recent progress in text-to-image (T2I) diffusion models (DMs) has enabled high-quality visual synthesis from diverse textual prompts. Yet, most existing T2I DMs, even those equipped with large language model (LLM)-based text encoders, remain text-pixel mappers -- they employ LLMs merely as text encoders, without leveraging their inherent reasoning capabilities to infer what should be visually depicted given the textual prompt. To move beyond such literal generation, we propose the think-then-generate (T2G) paradigm, where the LLM-based text encoder is encouraged to reason about and rewrite raw user prompts; the states of the rewritten prompts then serve as diffusion conditioning. To achieve this, we first activate the think-then-rewrite pattern of the LLM encoder with a lightweight supervised fine-tuning process. Subsequently, the LLM encoder and diffusion backbone are co-optimized to ensure faithful reasoning about the context and accurate rendering of the semantics via Dual-GRPO. In particular, the text encoder is reinforced using image-grounded rewards to infer and recall world knowledge, while the diffusion backbone is pushed to produce semantically consistent and visually coherent images. Experiments show substantial improvements in factual consistency, semantic alignment, and visual realism across reasoning-based image generation and editing benchmarks, achieving 0.79 on WISE score, nearly on par with GPT-4. Our results constitute a promising step toward next-generation unified models with reasoning, expression, and demonstration capacities.

Deep Learning · Large Language Models

Liew Yee Zhing, Andrew Tan, Anwar Majeed

Traditional hallucination detection fails on "Stubborn Hallucinations"—errors where LLMs are confidently wrong. We propose a geometric solution: Embedding-Perturbed Gradient Sensitivity (EPGS). We hypothesize that while robust facts reside in flat minima, stubborn hallucinations sit in sharp minima, supported by brittle memorization. EPGS detects this sharpness by perturbing input embeddings with Gaussian noise and measuring the resulting spike in gradient magnitude. This acts as an efficient proxy for the Hessian spectrum, differentiating stable knowledge from unstable memorization. Our experiments show that EPGS significantly outperforms entropy-based and representation-based baselines, providing a robust signal for detecting high-confidence factual errors.