Fine-tuning Done Right in Model Editing

Fine-tuning, a foundational method for adapting large language models, has long been considered ineffective for model editing. Here, we challenge this belief, arguing that the reported failure arises not from the inherent limitation of fine-tuning itself, but from adapting it to the sequential nature of the editing task, a single-pass depth-first pipeline that optimizes each sample to convergence before moving on. While intuitive, this depth-first pipeline coupled with sample-wise updating over-optimizes each edit and induces interference across edits. Our controlled experiments reveal that simply restoring fine-tuning to the standard breadth-first (i.e., epoch-based) pipeline with mini-batch optimization substantially improves its effectiveness for model editing. Moreover, fine-tuning in editing also suffers from suboptimal tuning parameter locations inherited from prior methods. Through systematic analysis of tuning locations, we derive LocFT-BF, a simple and effective localized editing method built on the restored fine-tuning framework. Extensive experiments across diverse LLMs and datasets demonstrate that LocFT-BF outperforms state-of-the-art methods by large margins. Notably, to our knowledge, it is the first to sustain 100K edits and 72B-parameter models,10 x beyond prior practice, without sacrificing general capabilities. By clarifying a long-standing misconception and introducing a principled localized tuning strategy, we advance fine-tuning from an underestimated baseline to a leading method for model editing, establishing a solid foundation for future research.

Does Generative Retrieval Overcome the Limitations of Dense Retrieval?

Generative retrieval (GR) has emerged as a new paradigm in neural information retrieval, offering an alternative to dense retrieval (DR) by directly generating identifiers of relevant documents. In this paper, we theoretically and empirically investigate how GR fundamentally diverges from DR in both learning objectives and representational capacity. GR performs globally normalized maximum-likelihood optimization and encodes corpus and relevance information directly in the model parameters, whereas DR adopts locally normalized objectives and represents the corpus with external embeddings before computing similarity via a bilinear interaction. Our analysis suggests that, under scaling, GR can overcome the inherent limitations of DR, yielding two major benefits. First, with larger corpora, GR avoids the sharp performance degradation caused by the optimization drift induced by DR's local normalization. Second, with larger models, GR's representational capacity scales with parameter size, unconstrained by the global low-rank structure that limits DR. We validate these theoretical insights through controlled experiments on the Natural Questions and MS MARCO datasets, across varying negative sampling strategies, embedding dimensions, and model scales. But despite its theoretical advantages, GR does not universally outperform DR in practice. We outline directions to bridge the gap between GR's theoretical potential and practical performance, providing guidance for future research in scalable and robust generative retrieval.

Speculative Safety-Aware Decoding

Despite extensive efforts to align Large Language Models (LLMs) with human values and safety rules, jailbreak attacks that exploit certain vulnerabilities continuously emerge, highlighting the need to strengthen existing LLMs with additional safety properties to defend against these attacks. However, tuning large models has become increasingly resource-intensive and may have difficulty ensuring consistent performance. We introduce Speculative Safety-Aware Decoding (SSD), a lightweight decoding-time approach that equips LLMs with the desired safety property while accelerating inference. We assume that there exists a small language model that possesses this desired property. SSD integrates speculative sampling during decoding and leverages the match ratio between the small and composite models to quantify jailbreak risks. This enables SSD to dynamically switch between decoding schemes to prioritize utility or safety, to handle the challenge of different model capacities. The output token is then sampled from a new distribution that combines the distributions of the original and the small models. Experimental results show that SSD successfully equips the large model with the desired safety property, and also allows the model to remain helpful to benign queries. Furthermore, SSD accelerates the inference time, thanks to the speculative sampling design.

Stop Spinning Wheels: Mitigating LLM Overthinking via Mining Patterns for Early Reasoning Exit

Large language models (LLMs) enhance complex reasoning tasks by scaling the individual thinking process. However, prior work shows that overthinking can degrade overall performance. Motivated by observed patterns in thinking length and content length, we categorize reasoning into three stages: insufficient exploration stage, compensatory reasoning stage, and reasoning convergence stage. Typically, LLMs produce correct answers in the compensatory reasoning stage, whereas reasoning convergence often triggers overthinking, causing increased resource usage or even infinite loops. Therefore, mitigating overthinking hinges on detecting the end of the compensatory reasoning stage, defined as the Reasoning Completion Point (RCP). RCP typically appears at the end of the first complete reasoning cycle and can be identified by querying the LLM sentence by sentence or monitoring the probability of an end-of-thinking token (e.g., \texttt{</think>}), though these methods lack an efficient and precise balance. To improve this, we mine more sensitive and consistent RCP patterns and develop a lightweight thresholding strategy based on heuristic rules. Experimental evaluations on benchmarks (AIME24, AIME25, GPQA-D) demonstrate that the proposed method reduces token consumption while preserving or enhancing reasoning accuracy.

How Do LLM-Generated Texts Impact Term-Based Retrieval Models?

As more content generated by large language models (LLMs) floods into the Internet, information retrieval (IR) systems now face the challenge of distinguishing and handling a blend of human-authored and machine-generated texts. Recent studies suggest that neural retrievers may exhibit a preferential inclination toward LLM-generated content, while classic term-based retrievers like BM25 tend to favor human-written documents. This paper investigates the influence of LLM-generated content on term-based retrieval models, which are valued for their efficiency and robust generalization across domains. Our linguistic analysis reveals that LLM-generated texts exhibit smoother high-frequency and steeper low-frequency Zipf slopes, higher term specificity, and greater document-level diversity. These traits are aligned with LLMs being trained to optimize reader experience through diverse and precise expressions. Our study further explores whether term-based retrieval models demonstrate source bias, concluding that these models prioritize documents whose term distributions closely correspond to those of the queries, rather than displaying an inherent source bias. This work provides a foundation for understanding and addressing potential biases in term-based IR systems managing mixed-source content.

LLM4MEA: Data-free Model Extraction Attacks on Sequential Recommenders via Large Language Models

Recent studies have demonstrated the vulnerability of sequential recommender systems to Model Extraction Attacks (MEAs). MEAs collect responses from recommender systems to replicate their functionality, enabling unauthorized deployments and posing critical privacy and security risks. Black-box attacks in prior MEAs are ineffective at exposing recommender system vulnerabilities due to random sampling in data selection, which leads to misaligned synthetic and real-world distributions. To overcome this limitation, we propose LLM4MEA, a novel model extraction method that leverages Large Language Models (LLMs) as human-like rankers to generate data. It generates data through interactions between the LLM ranker and target recommender system. In each interaction, the LLM ranker analyzes historical interactions to understand user behavior, and selects items from recommendations with consistent preferences to extend the interaction history, which serves as training data for MEA. Extensive experiments demonstrate that LLM4MEA significantly outperforms existing approaches in data quality and attack performance, reducing the divergence between synthetic and real-world data by up to 64.98% and improving MEA performance by 44.82% on average. From a defensive perspective, we propose a simple yet effective defense strategy and identify key hyperparameters of recommender systems that can mitigate the risk of MEAs.

A Survey of Link Prediction in N-ary Knowledge Graphs

N-ary Knowledge Graphs (NKGs) are a specialized type of knowledge graph designed to efficiently represent complex real-world facts. Unlike traditional knowledge graphs, where a fact typically involves two entities, NKGs can capture n-ary facts containing more than two entities. Link prediction in NKGs aims to predict missing elements within these n-ary facts, which is essential for completing NKGs and improving the performance of downstream applications. This task has recently gained significant attention. In this paper, we present the first comprehensive survey of link prediction in NKGs, providing an overview of the field, systematically categorizing existing methods, and analyzing their performance and application scenarios. We also outline promising directions for future research.

KnowCoder-V2: Deep Knowledge Analysis

Deep knowledge analysis tasks always involve the systematic extraction and association of knowledge from large volumes of data, followed by logical reasoning to discover insights. However, to solve such complex tasks, existing deep research frameworks face three major challenges: 1) They lack systematic organization and management of knowledge; 2) They operate purely online, making it inefficient for tasks that rely on shared and large-scale knowledge; 3) They cannot perform complex knowledge computation, limiting their abilities to produce insightful analytical results. Motivated by these, in this paper, we propose a \textbf{K}nowledgeable \textbf{D}eep \textbf{R}esearch (\textbf{KDR}) framework that empowers deep research with deep knowledge analysis capability. Specifically, it introduces an independent knowledge organization phase to preprocess large-scale, domain-relevant data into systematic knowledge offline. Based on this knowledge, it extends deep research with an additional kind of reasoning steps that perform complex knowledge computation in an online manner. To enhance the abilities of LLMs to solve knowledge analysis tasks in the above framework, we further introduce \textbf{\KCII}, an LLM that bridges knowledge organization and reasoning via unified code generation. For knowledge organization, it generates instantiation code for predefined classes, transforming data into knowledge objects. For knowledge computation, it generates analysis code and executes on the above knowledge objects to obtain deep analysis results. Experimental results on more than thirty datasets across six knowledge analysis tasks demonstrate the effectiveness of \KCII. Moreover, when integrated into the KDR framework, \KCII can generate high-quality reports with insightful analytical results compared to the mainstream deep research framework.

On the Scaling of Robustness and Effectiveness in Dense Retrieval

Robustness and Effectiveness are critical aspects of developing dense retrieval models for real-world applications. It is known that there is a trade-off between the two. Recent work has addressed scaling laws of effectiveness in dense retrieval, revealing a power-law relationship between effectiveness and the size of models and data. Does robustness follow scaling laws too? If so, can scaling improve both robustness and effectiveness together, or do they remain locked in a trade-off? To answer these questions, we conduct a comprehensive experimental study. We find that:(i) Robustness, including out-of-distribution and adversarial robustness, also follows a scaling law.(ii) Robustness and effectiveness exhibit different scaling patterns, leading to significant resource costs when jointly improving both. Given these findings, we shift to the third factor that affects model performance, namely the optimization strategy, beyond the model size and data size. We find that: (i) By fitting different optimization strategies, the joint performance of robustness and effectiveness traces out a Pareto frontier. (ii) When the optimization strategy strays from Pareto efficiency, the joint performance scales in a sub-optimal direction. (iii) By adjusting the optimization weights to fit the Pareto efficiency, we can achieve Pareto training, where the scaling of joint performance becomes most efficient. Even without requiring additional resources, Pareto training is comparable to the performance of scaling resources several times under optimization strategies that overly prioritize either robustness or effectiveness. Finally, we demonstrate that our findings can help deploy dense retrieval models in real-world applications that scale efficiently and are balanced for robustness and effectiveness.

Inference-time Alignment in Continuous Space

Aligning large language models with human feedback at inference time has received increasing attention due to its flexibility. Existing methods rely on generating multiple responses from the base policy for search using a reward model, which can be considered as searching in a discrete response space. However, these methods struggle to explore informative candidates when the base policy is weak or the candidate set is small, resulting in limited effectiveness. In this paper, to address this problem, we propose Simple Energy Adaptation ($\textbf{SEA}$), a simple yet effective algorithm for inference-time alignment. In contrast to expensive search over the discrete space, SEA directly adapts original responses from the base policy toward the optimal one via gradient-based sampling in continuous latent space. Specifically, SEA formulates inference as an iterative optimization procedure on an energy function over actions in the continuous space defined by the optimal policy, enabling simple and effective alignment. For instance, despite its simplicity, SEA outperforms the second-best baseline with a relative improvement of up to $ \textbf{77.51%}$ on AdvBench and $\textbf{16.36%}$ on MATH. Our code is publicly available at https://github.com/yuanyige/SEA