Papers

The continuous evolution of large language models drives escalating demands on data scale and quality, and as different training stages impose increasingly tailored data requirements, systematic organization of high-quality corpora becomes indispensable. Existing corpus construction pipelines confine the resulting corpora to flat, undifferentiated document collections, universally lacking systematic knowledge organization. We present Cortex, to our knowledge the first framework that elevates web-scale corpus construction from flat document filtering to structured knowledge organization through an Ontological Corpus Graph (OCG), a three-layer heterogeneous structure unifying a quality-refined content layer, a hierarchical lightweight ontology layer via LLM-driven automated evolution, and a cross-domain alignment layer enabling inter-domain association at arbitrary taxonomic resolution. Comprehensive experiments confirm the effectiveness of Cortex. In particular, we leverage the OCG to synthesize CortexBench, a cross-domain search-and-reasoning benchmark whose evaluation across eight frontier LLMs validates the effectiveness of quality refinement, domain organization, and cross-domain data synthesis. We will publicly release the complete codebase, a 24.14B-token refined corpus with its OCG, and CortexBench. The data is available at https://github.com/zjukg/CORTEX .

Contextual language models conflate grammatical gender and social semantic bias in gendered languages such as Spanish. Existing gender debiasing approaches only operate on static word embeddings leaving contextual representations unexplored for this two dimensional gender disentanglement. To address the this issue, we make the first attempt to disentangle grammatical gender from semantic contamination for contextual embeddings. We construct both controlled templates and natural Wikipedia contexts to build balanced datasets of inanimate nouns, and design a framework equipped with centroid, Support Vector Machine (SVM) and Linear Discriminant Analysis (LDA) gender direction estimators as well as contamination-aware weighting strategies. A set of dual-objective evaluation metrics is proposed to balance the suppression of grammatical gender leakage on inanimate nouns and the preservation of semantic gender distinctions for occupation terms. The results reveal that unweighted controlled contexts yield the purest grammatical gender direction, and the centroid estimator achieves better performance than discriminative baselines.

Recent breakthroughs in foundation models and Large Language Models (LLMs) have introduced new opportunities for studying and decoding genomic sequences. Several state-of-the-art approaches, such as DNABERT2, rely on transformer-based architectures, while others, such as ConvNova, still build upon more conventional convolutional models. However, systematic benchmark comparisons across these methods remain scarce. Given that transformer-based models require extensive and costly pretraining, it is crucial to evaluate whether their performance gains justify this overhead. Moreover, LLMs such as DNABERT2 typically rely on Byte Pair Encoding (BPE) tokenization, whose relevance for DNA sequence representation is still debated within the genomics community. In this work, we investigate three key questions: (i) do transformer-based models provide sufficient improvements on fine-tuning tasks upon heavy pretraining, (ii) what is the actual contribution of pretraining in this setting, and (iii) how does BPE tokenization impact performance on genomics-related tasks?

Chain-of-thought (CoT) prompting improves LLM reasoning, but the source is contested: do the intermediate steps help because they carry useful semantic content, or because conditioning on more tokens buys extra computation before the model commits to an answer? We bring two lines of evidence to bear. First, in distribution: we repeatedly sample each model on the same question and pair a shorter with a longer of its own natural generations that follow the same reasoning plan, so nothing is rewritten and both traces are genuinely in-distribution. Across 25 models the extra tokens leave accuracy essentially unchanged for every independently-trained reasoner, and a blind analysis of the surplus tokens shows that what gain exists elsewhere tracks validation- and checking-content, not verbosity per se. Second, as a controlled intervention, we ask whether two traces expressing the same semantic content (the same facts, operations, and intermediate values, verified through directed acyclic graph equivalence) produce different outcomes when one is more verbose, using a dual-validator design across four targets and eight benchmarks with number-redacted completion and stratified bootstrap confidence intervals. Verbose traces do improve accuracy (25 of 32 benchmark-target cells are positive under at least one validator), but the effects are modest (typically 1-4 points) and depend on the quality of the verbose prose, not merely its length. Under maximum numerical redaction the effect is amplified (median 3.24x across four arithmetic benchmarks), and length-matched non-reasoning filler recovers none of it. Both lines converge: what matters is what the extra tokens do (the reasoning and validation content they carry), not how many there are, a picture neither a pure forward-pass-compute nor a pure semantic-content account fully explains.

Quantifying how meaning propagates through communicative exchanges remains underdeveloped in computational linguistics. Here we introduce an information-theoretic framework that quantifies the directed flow of semantic content between interlocutors and decomposes multi-source contributions into redundant, unique, and synergistic components. Our approach leverages large language models as probabilistic estimators of natural language to compute two measures: semantic transfer entropy (STE), which captures directed predictive influence between speakers, and semantic partial information decomposition (SPID), which resolves how multiple sources jointly shape a target's language. Across four experiments we show that the framework detects reduced information flow in cognitively rigid dialogue, captures the dominant role of persuaders in shaping discourse, distinguishes high- from low-quality psychotherapy by the directionality of therapist-client information exchange, and reveals synergistic premise contributions in argumentative essays. This framework opens new avenues for studying information dynamics in digital discourse, pedagogical interactions, clinical dialogues, and any domain in which the structure of linguistic exchange is of research relevance.

Retrieval-Augmented Generation (RAG) mitigates hallucinations in Large Language Models (LLMs) by grounding the generation process on external knowledge. However, standard RAG approaches struggle with multi-hop reasoning. While recent graph-based RAG methods improve the retrieval of interconnected chunks, they often rely on computationally expensive and error-prone LLM-based extraction pipelines. To address these issues, we propose TIGRAG (Token-Induced GraphRAG), an efficient graph-augmented RAG framework based on a token co-occurrence Knowledge Graph. TIGRAG directly models topological relationships between tokens using sliding-window co-occurrence statistics, thus enabling scalable graph construction. During inference, it combines graph-based semantic expansion and neural reranking to retrieve interconnected evidence for multi-hop reasoning. Specifically, it introduces an iterative entity-driven retrieval strategy that progressively expands the query using bridging entities extracted from previously retrieved contexts. We evaluated TIGRAG on three widely adopted multi-hop Question Answering (QA) benchmarks. Experimental results demonstrated that our framework consistently outperforms dense retrieval and graph-based RAG methods in both retrieval and downstream QA tasks, while substantially reducing indexing time, inference latency, and prompt footprint.

Large-language models have proven to be remarkable if inconsistent parrots of public attitudes and opinions. The extent to which LLMs are able to produce reasonable approximations of cultural taste remains an open empirical question that becomes more urgent by the day, with market research companies already offering provisional 'synthetic' survey panels and the contamination of standard survey data from LLM-generated responses. In this study, we build on past work on silicon sampling by extending considerations of their ecological, relational, and positional fidelity in the doomain of cultural tastes. We use large-language models from OpenAI, Anthropic, and DeepSeek to produce 554,940 silicon surrogates of survey respondents from the Survey of Public Participation in the Arts (SPPA). We find these silicon surrogates' tastes to be highly stylized facsimiles of human tastes. First, silicon samples are super-omnivorous with a systematic postive-bias for liking. These individual-level bias of silicon samples are not well-explained by the WEIRD-bias often discussed in the literature. Second, the complex relationality in real taste structures is completely distorted among silicon samples. Third, very little of the known cultural alignment between tastes and social space are preserved. Silicon samples juvenilize age-taste associations, resurrect anachronistic class-taste associations, and caricaturize gender- and race-taste associations. Key words: AI, taste, consumption, culture, silicon sampling, meta-analysis.

While large language models have been dominating the research landscape recently, small language models remain highly relevant across various domains; yet, they receive far less attention. In this study, we investigate how smaller language models perform during the generation stage within a Retrieval-Augmented Generation (RAG) system. To benchmark these models effectively, we utilised both open-source and proprietary datasets covering diverse subject areas and question types. Our findings demonstrate that a RAG system with small language models can be executed directly on-device without requiring any GPU hardware within a reasonable time. The experimental code and links to the supplementary materials can be accessed through the GitHub repository: https://github.com/SibNN/SLM-RAG-EVAL.

Since intelligence fundamentally relies on efficient skill acquisition (Chollet, 2019), the ability to leverage skills is critical. For LLMs, skills, manually authored or extracted from task trajectories, are textual recipes encoding mature problem-solving experience and are critical to agentic capabilities. Despite widespread deployment, their utility is limited by the model's ability to comprehend and follow skill instructions, especially under complex and long-context scenarios, where key instructions are difficult to locate and adhere to. To address this limitation, we propose ParametricSkills, a framework that can convert free-form textual skills into parameters at test time, enabling context-free skill exploitation. Specifically, we first construct a large-scale, high-quality skill library, and synthesize single-turn and multi-turn skill exploitation trajectories built around these skills with OpenCode. Using these data, we then train a hypernetwork that parameterizes both the skill content and the test-time exploitation methodology by receiving textual skills and converting them into LoRA adapters. Experimental results on six complex software engineering (SWE) subtasks demonstrate that, the proposed ParametricSkills averagely outperforms in-context learning by 6.44 points as judged by DeepSeek-V4-Flash, while also achieving significantly higher BERT Score and F1 score, confirming its effectiveness. Beyond performance, we further find that parametric skills, being inherently accumulative, offer a preliminary yet promising avenue toward test-time continual learning.

Graph anomaly detection (GAD) on text-attributed graphs (TAGs) is vital for applications such as fraud detection and academic integrity verification. Existing approaches generally fall into two paradigms. GNN-based methods effectively capture structural patterns but struggle to capture fine-grained textual semantics. Methods integrating LLMs with graphs improve semantic understanding yet fail to fully comprehend topological relationships among neighboring nodes. Moreover, both paradigms overlook the correspondence between textual semantics and graph topological relationships, limiting their ability to identify nodes whose semantics are inconsistent with their neighborhoods. In this paper, we formalize TAG anomaly detection as a node-to-neighborhood semantic consistency problem, where anomalies may arise from either textual semantic mismatch or topological deviation between a node and its neighbors. We propose N2NSC (Node-to-Neighborhood Semantic Consistency), a framework that captures the correspondence between graph topology and textual semantics through two complementary fusion paths. The two pathways work synergistically, enabling the LLM to fully leverage both textual and structural neighborhood information for anomaly detection. Extensive experiments across eight datasets demonstrate that N2NSC consistently outperforms current state-of-the-art methods.

Long-horizon tool agents are bottlenecked by how their context grows toward the limits of the context window. Recent systems make context management agent- or system-controlled, but they either learn compression policies that discard evidence or manage context in a layer the agent never sees. We argue that both miss a more basic gap: frontier language models are proprioceptively blind to their own context. From the prompt alone they cannot reliably infer block size, recency, or the remaining budget, all of which are needed for keep-or-archive decisions. We introduce VISTA (Visible Internal State for Tool Agents), a training-free, model-agnostic layer that represents working memory as typed addressable blocks, surfaces a runtime dashboard of token usage, recency, archive status, and remaining budget, and archives blocks as recoverable full-fidelity payloads. On LOCA-Bench, BrowseComp-Plus, and GAIA, the same untrained interface transfers across 1M-, 100K-, and 10K-scale trajectories. On LOCA-Bench it lifts Gemini-3-Flash from 22.7 to 50.7%, reaches 58.0% on BrowseComp-Plus, and remains competitive on GAIA. Gains grow with context pressure and transfer across backbones, while ablations confirm that the dashboard matters beyond archive and recovery tools.

Diversity in LLM mathematical reasoning is critical for exploration, but common diversity metrics mostly capture surface-level variation rather than differences in how a problem is solved. We address this gap by introducing approach-level diversity: variation in strategies across correct solutions to the same problem. Using a human-calibrated LLM judge framework, we show that prior diversity measures are unreliable proxies for approach-level diversity, and this mismatch carries over to diversity-aware RLVR, where target metrics are preserved while approach-level diversity declines. Investigating when approach-level diversity helps and whether it can be directly induced, we find that approach-diverse candidate sets improve test-time scaling. However, optimizing an LLM judge diversity reward during training causes the policy to exploit judge-specific preferences rather than broaden its approaches, leaving direct optimization of approach-level diversity as an open problem. Together, our work introduces the notion of approach-level diversity and uncovers a systematic divergence between surface- and approach-level signals, marking a step toward LLMs that reason in genuinely diverse, human-like ways.

Large Language Models (LLMs) often fail to maintain instruction hierarchies (IH) when processing multi-source inputs with varying role-level priorities, paradoxically adhering to lower-priority directives during conflicts. While existing defenses mitigate this issue, they are largely restricted to single-turn scenarios and require expensive fine-tuning. In this paper, we formalize this failure mode in multi-turn contexts via a Jensen-Shannon Divergence (JSD) framework, uncovering a pervasive role-influence inversion phenomenon where subordinate inputs override superior roles. To rectify this without training, we propose IHDec (Instruction Hierarchy-steered Decoding). IHDec leverages JSD to automatically detect token-level hierarchy violations and dynamically executes contrastive decoding to suppress misaligned subordinate roles. Extensive evaluations demonstrate that IHDec outperforms training-based baselines in multi-turn conflicts while fully preserving general response quality. Furthermore, IHDec strengthens safety against adversarial prompt injections and exhibits a robust scaling synergy with larger models. The Code is available at https://github.com/nxcolelxu/IHDec.git

Retrieval-augmented generation (RAG) typically retrieves a fixed number of passages for every query. This is wasteful when the reader already knows the answer, and it can be harmful when irrelevant or partially relevant passages distract the reader. We formulate adaptive RAG as calibrated retrieval-budget allocation: given a query, decide whether to answer closed-book, retrieve a compact context (k=1), retrieve a full context (k=5), or abstain. The contribution is a probability interface rather than a new raw uncertainty signal. We calibrate sequence log-probability and prefix-logit uncertainty signals into probabilities of correctness, then use these probabilities for graded context selection, selective abstention, and explicit latency/token trade-offs. Across core QA experiments on TriviaQA, Natural Questions, and MS MARCO, with auxiliary PopQA motivation and Qwen/Llama family checks, diagnostic out-of-fold calibration improves probability quality dramatically: for sequence log-probability, ECE drops from 0.275 to 0.062 on TriviaQA, 0.643 to 0.009 on NQ, and 0.711 to 0.031 on MS MARCO. Graded retrieval improves full-context and passage-budget frontiers for both our signal and TARG-style prefix entropy/margin, while retrieval-call AUC remains essentially tied with binary gating because k=1 is still a retrieval call. Held-out train/validation/test threshold experiments report deployable operating points. At matched-accuracy frontier operating points, a measured cost model reveals that gating is not universally faster: it increases latency by about 27% on Qwen3-8B but saves about 8% on Qwen3-32B. These results support a nuanced view of adaptive RAG: calibrated confidence is best understood as a reusable interface for allocating retrieval budget under task and system constraints.

Most coding-agent benchmarks are static: an agent receives a complete task description up front and is judged only by its final code. Real coding assistance is interactive, with users clarifying goals, adding constraints, and correcting mistakes over multiple turns. We introduce SWE-Together, a multi-turn benchmark reconstructed from real user-agent coding sessions. To make real interactions verifiable, we curate 109 repository-level tasks from 11,260 recorded sessions, selecting sessions with recoverable repository states, clear user goals, and observable outcomes. To replay these interactions across agents, we build a reactive LLM-based user simulator that preserves the original users' intents and provides feedback when the coding agent's progress requires it. To evaluate agents as collaborators, we measure both final repository correctness and the number of corrective feedback turns required during the interaction. Experiments with frontier coding agents show that stronger agents generally achieve higher final success rates while requiring fewer interventions, suggesting an improved user experience.

Reinforcement learning with verifiable rewards (RLVR) is bottlenecked by hard prompts on which correct trajectories have low probability, so sampling misses them within a practical budget and leaves the policy update with little useful signal. We frame such zero-hit prompts as RLVR's sampling frontier, where new reasoning behavior is most valuable yet least likely to be sampled. Importantly, failed rollouts can be informative: they expose where the model's reasoning went wrong. We introduce LatentRevise, a first-order latent revision method that recovers training signal for this zero-hit regime. Given a failed rollout and the gold answer as an anchor, LatentRevise optimizes the input embeddings of its reasoning prefix under two complementary gradients, moving the prefix away from the failed continuation and toward the gold answer. The optimization is constrained to the convex hull of the model's vocabulary embeddings, so each update moves the latent toward a real token embedding rather than an arbitrary feature direction. We find that continuations from the revised prefix lengthen, exhibit self-reflection, and reach correct answers missed by the original rollouts. Used as training data, these trajectories improve SFT and RLVR on math benchmarks over standard baselines.

The alignment of language models is typically studied through the lens of capability benchmarks, but the dynamics of how models change during post-training remain poorly understood. We argue that the physical sciences, and thermodynamic phase-transition theory in particular, offer a principled and underexplored vocabulary for reasoning about these dynamics. As a case study, we instantiate this position through the lens of material Crystallization, which is a well-studied thermodynamic phase transition. For tasks like random number generation, this breaks into 3 phases: (1) the high entropy liquid phase in the pretrained model, with many distinct sampling distributions promptable from the model; (2) the nucleation phase caused by supervised finetuning, in which behavior collapses onto a single seed distribution present in the pretrained LLM; and (3) a settling phase in which reinforcement learning techniques redistribute probability of the collapsed distribution, but largely keep it concentrated on the same options as the seed distribution. We propose intuitive metrics to verify the transitions between these phases, and validate the idea across a range of random tasks. Crystallization is one instance of a broader class of physical frameworks we believe alignment research should import to answer questions about where alignment-induced structure comes from, why it converges where it does, and what it fundamentally cannot change.

Rubric-based scoring has become a widely used paradigm in model evaluation, typically with LLM-as-a-Judge (LaaJ) for rubric scoring. However, the reliability of LaaJ for rubric scoring remains underexplored. This concern is especially pronounced in agentic scenarios, where long, complex outputs further challenge reliable scoring. To address this, we conduct a systematic meta-evaluation of LaaJ reliability for rubric verification. We introduce RuVerBench, the first benchmark for assessing LaaJ reliability in rubric verification for agentic scenarios. RuVerBench covers two prevalent agentic domains, deep research and agentic coding, with 2,458 instances, each containing a model-generated output, a rubric, and a human-annotated label indicating whether the output satisfies the rubric. Using RuVerBench, we evaluate numerous frontier LLMs and find that even the most advanced models achieve strong performance but still exhibit substantial noise. We further analyze the impact of key LaaJ strategies, including prompt design, batching, and majority voting, on rubric verification. We find that weaker models are more sensitive to prompt variations, batched verification presents a trade-off between accuracy and efficiency, and majority voting yields effective but diminishing returns. We have released our dataset and code to facilitate future research: https://github.com/THU-KEG/RuVerBench.

Agent memory systems are increasingly evaluated against RAG and full-context baselines, but reported gains often mix changes in the memory method with changes in the language model, embedding model, or retrieval pipeline, making it unclear what is actually being measured. We present MemDelta, a controlled evaluation protocol that varies one component at a time on LongMemEval-S (500 questions, 50+ sessions, three model families). Four findings emerge: (1) verbatim RAG matches full-context GPT-4o-mini (47.2% vs. 49.8%, p = 0.34), but the ranking reverses across models: Gemini gains +14pp from full context, while Sonnet gains +31pp from RAG, partly because it refuses 63% of full-context queries; (2) swapping only the embedding model in an identical pipeline shifts accuracy by +6.2pp at n = 500 (p = 0.004), and Mem0 beats MiniLM-RAG by +11pp but loses to cloud-RAG by 1.2pp, so one variable flips the conclusion; (3) agent self-memory (42%) underperforms basic retrieval (47%); (4) on 2 of 6 question types (n = 88), Mem0 matches cloud RAG (72.7% vs. 73.9%, p = 1.0) at 50x the cost, suggesting narrow rather than general gains. We recommend memory evaluations fix embedding models across comparisons, stratify by model family, and report write-path cost before attributing gains to architecture.

This study demonstrates an alignment of per-word processing time in a popular state-space language model Mamba and human readers. In Mamba, the recurrent state transition at each layer conceptually takes some duration of time, the discretization timestep $\Delta_t$, determined dynamically in response to the input. Using a naturalistic reading dataset, we show that the per-word timestep from Mamba is a significant predictor of human reading times, and remains significant even when known predictors such as GPT-2 surprisal are controlled for. We further suggest, through formal analysis of Mamba's architecture and internal dynamics, that Mamba can serve as a new, valuable lens to look at human real-time language processing with ever-updated memory, because it allows us to look at how each module (layer) weighs short- and long-term information retention, and how noise may interact with dynamic, continuous memory representation. Code is available online.

As agentic AI systems tackle more complex mathematical tasks, they increasingly rely on information retrieval (IR) to search problem databases, theorem libraries, and educational resources. However, choosing the right retriever remains difficult, as it is infeasible to directly isolate its effect on downstream performance. On the other hand, existing retrieval-specific benchmarks often fail to capture fine-grained mathematical relevance, penalizing relevant documents. We address this gap by introducing SABER-Math, the first fully automated benchmark for evaluating mathematical IR without expert annotation. Starting from 283K high-school-level math problems with solutions, SABER-Math builds challenging reranking tasks in three steps: (i) first, LLMs extract concise solution summaries and mathematical topics for each problem; (ii) then, per-query relevant documents are discovered using ontology topic-based and lexical solutions-summary-based similarities, and (iii) finally, a Swiss-style LLM preference tournament produces fine-grained relevance ratings for the documents. We evaluate lexical retrievers, specialized mathematical retrieval systems, and recent embedding models. We find that while modern embedding models substantially outperform classical and math-specific baselines, even the strongest systems struggle in symbol-heavy domains like Algebra and Calculus. Importantly, we show that general-purpose IR benchmarks such as MTEB do not reliably predict mathematical performance, especially for recent embedding models, highlighting the need for math-specific retrieval benchmarks.

Modern large language models (LLMs) reach 60-70% diagnostic accuracy on complex clinical case benchmarks, but accuracy alone cannot distinguish stable clinically-grounded reasoning from pattern matching. We introduce clinical reasoning graphs, structured graph representations extracted from free-text LLM diagnostic traces using a domain-grounded ontology with 5 node types and 7 edge types. We apply this pipeline to 750 traces from five LLMs across 50 New England Journal of Medicine Clinicopathological Conference cases and three prompt conditions, and test whether diagnostic traces show stable structured reasoning patterns, or diagnostic schemas, for clinically similar cases. We operationalize this as higher graph similarity among clinically similar cases than among clinically dissimilar ones. Across 15 model-condition comparisons, within-cluster and between-cluster composite similarity are nearly equal, and no comparison survives multiple-testing correction; a component-level analysis finds any residual content signal far below schema scale. Graph similarity is also nearly identical for pairs of models that are both correct (0.488) and both incorrect (0.484), suggesting that graph structure captures a dimension not reflected in diagnostic accuracy. Structured reflection prompting increases explicit discriminating-feature analysis within traces (+33%) but does not increase cross-case consistency. These results show diagnostic competence without schema-scale reasoning consistency, and indicate that final-answer accuracy should be complemented by process-level evaluation. We release the ontology, extraction pipeline, validation protocol, and the extracted reasoning graphs and similarity artifacts as resources for structured evaluation of LLM clinical reasoning.

This study analyzes the novelty distribution of scholarly papers in the field of Library and Information Science (LIS) in China, with a focus on differences across journals, research topics, and time periods. Articles published in Chinese LIS journals indexed by the Chinese Social Sciences Citation Index (CSSCI) from 2000 to 2022 were collected as the research sample. BERTopic was applied to paper abstracts to identify research topics, and novelty scores were calculated based on the combinatorial innovation theory of reference pairs cited by focal papers. The study then examined the novelty of papers under different topics and further analyzed author collaboration patterns to explain how collaboration may be associated with paper novelty. The results show that archival research topics generally have lower novelty, whereas topics related to journal evaluation and patent technology display higher novelty in Chinese LIS research. Overall, the novelty of papers in this field has gradually increased over time. Papers with different topics and novelty levels also show distinct collaboration patterns: low-novelty topics are more often associated with solo authorship, while high-novelty topics tend to involve a higher proportion of inter-institutional collaboration. This study reveals the topic-level characteristics and temporal trends of novelty in Chinese LIS research and provides a new perspective for understanding how research topics and collaboration patterns influence scholarly innovation.

Knowledge distillation (KD) is a key technique for compressing Large Language Models (LLMs), yet methods relying on a single KL objective often fail to balance primary distribution fitting with long-tail probability modeling, limiting both generation quality and generalization. To address this, we analyze the complementary roles of forward and reverse KL divergence (FKL/RKL) in distribution alignment from theoretical and empirical perspectives. We then propose a reinforcement-learning-based adaptive KL-weighted distillation framework, in which a policy network dynamically assigns weights to FKL and RKL based on teacher-student distributional characteristics, guided by immediate reward signals to achieve dual alignment on principal and long-tail modes. Extensive experiments demonstrate consistent improvements across Rouge-L and BertScore metrics, surpassing greedy heuristics by 0.4-0.6 points and outperforming other baseline methods on diverse benchmarks.

Agentic search equips large language models with dynamic retrieval abilities, but existing reinforcement learning methods remain limited by reward sparsity in knowledge boundary calibration – deciding when to trust parametric memory, when to rely on retrieved evidence, and when to abstain. Binary rewards can penalize undesirable outcomes, but provide little guidance on the reasoning process required to make calibrated decisions across different knowledge states. To address this, we propose KbSD (Knowledge boundary Self-Distillation), a framework that tackles this limitation through dense token-level supervision, outcome-level sparse rewards, and quadrant-adaptive optimization. KbSD constructs a hint-augmented teacher, architecturally identical to the student, that receives explicit knowledge boundary signals – including parametric certainty, retrieval quality, and ground-truth answers – to generate calibrated reasoning demonstrations. This information-asymmetric self-distillation enables dense supervision without requiring a larger external model. To further account for the heterogeneous reasoning distributions across knowledge states, we introduce a quadrant-adaptive distillation objective: reverse KL for concentrated integration, forward KL for diverse refusal, and Pareto-optimal bidirectional KL for asymmetric quadrants requiring both precision and coverage. Experiments on multiple benchmarks show that KbSD consistently improves both task accuracy and hallucination mitigation over strong baselines, with the largest gains appearing in the challenging quadrants where sparse rewards are least informative.

With the rise of data-intensive science, algorithms have become central to scientific research. In academic papers, algorithms are mentioned for different purposes, such as describing, using, comparing, or improving methods for specific research tasks. Identifying these purposes can reveal relationships among algorithms and help assess their roles and value. Taking natural language processing (NLP) as an example, this study proposes a sentence-level framework for identifying, analyzing, and tracing the evolution of motivations for mentioning algorithms. We first identify algorithm entities and algorithm-related sentences from full-text papers through manual annotation and machine learning. We then classify mention motivations using pretrained models and data augmentation, and analyze their distribution and temporal evolution. The results show that deep learning models trained with augmented data outperform traditional machine learning models in motivation classification. In NLP papers, more than half of algorithm-related sentences express direct use, whereas improvement is the least frequent motivation. The diversity of motivations has increased over time. For specific algorithm categories, grammar-based algorithms are more often mentioned for description, while machine learning algorithms are more often mentioned for use. Over time, use motivations have gradually replaced description motivations across different algorithms, and the number of motivation types associated with individual algorithms has declined significantly. This study reveals how authors mention algorithm entities in academic writing and provides a basis for future research on algorithm relationship identification and algorithm impact evaluation.

Language model loss follows remarkably regular scaling laws over model and data size, yet it remains unclear why the aggregate loss should exhibit a power-law form. Existing explanations often attribute this regularity to a heavy-tailed spectrum of pattern difficulty in natural language, but this view has not been directly validated at token-level granularity in large-scale real-data training. We present a token-level framework that decomposes scaling laws into localized learning events of individual contextualized tokens. By fitting token loss trajectories with sigmoids, we show that token learning is concentrated in localized transitions, giving rise to a learning-time spectrum that dominates the scaling-law shape. Across more than one hundred pre-training runs on large and diverse real-language corpora with modern LLM architectures, scaling up to 6B parameters and 300B training tokens, the measured learning-time spectrum quantitatively reconstructs the validation loss derivative along the training-step $T$, data-scale $D$, and model-scale $M$ axes. We further show that the same signal is actionable: by reshaping the training distribution according to when tokens become learnable, we alter the optimization trajectory and achieve 11% faster validation-loss reduction. These results provide direct empirical evidence that scaling laws are governed primarily by the distribution of token-level learning times, and that this distribution can be used not only to explain scaling behavior but also to improve training performance.

The quadratic computational cost of traditional attention mechanisms poses a major bottleneck to the scalability and practical deployment of large language models (LLMs), particularly in long-context scenarios. To improve efficiency, existing approaches often enforce rigid structural constraints such as local attention windows. However, these strategies typically lead to substantial performance degradation on tasks requiring precise long-range recall. In this work, we propose MATCH, a scalable and efficient framework that augments sparsified attention mechanisms with dynamically integrated in-context information through an efficient retrieval system. Empirical results show that MATCH significantly improves the performance of sparse-attention models on both synthetic and real-world natural-language tasks. These findings highlight the versatility of MATCH as a general approach for enhancing in-context retrieval capabilities while maintaining the efficiency benefits of sparse attention architectures.

Most studies on technology development have been conducted from a thematic perspective, but the topics are coarse-grained and insufficient to accurately represent technology. The development of automatic entity recognition techniques makes it possible to extract technology-related entities on a large scale. Thus, we perform a more accurate analysis of technology development from an entity-centric perspective. To begin with, we extract technology-related entities such as methods, datasets, metrics, and tools in articles on Natural Language Processing (NLP), and we apply a semi-automatic approach to normalize the entities. Subsequently, we calculate the z-scores of entities based on their co-occurrence networks to measure their impact. We then analyze the development trends of new technologies in the NLP domain since the beginning of the 21st century. The findings of this paper include three aspects: Firstly, the continued increase in the average number of entities per paper implies a growing burden on researchers to acquire relevant technical background knowledge. However, the emergence of pre-trained language models has injected new vitality into the technological innovation of the NLP domain. Secondly, Methods dominate among the 179 high-impact entities. An analysis of the z-score trend about the top 10 entities reveals that pre-trained language models, exemplified by BERT and Transformer, have become mainstream in recent years. Unlike the trend of the other eight method entities, the impact of Wikipedia dataset and BLEU metric has continued to rise in the long term. Thirdly, in recent years, there has been a remarkable surge in popularity for new high-impact technologies than ever before, and their acceptance by researchers has accelerated at an unprecedented speed. Our study provides a new perspective on analyzing technology development in a specific domain.

While Large Language Models (LLMs) excel as static solvers, transforming them into autonomous agents remains challenging. This transition requires continuous environmental interaction, yet current agents lack the necessary persistent procedural memory. Existing approaches predominantly employ Retrieval-Augmented Generation (RAG) to inject explicit textual guidelines into model contexts. However, relying solely on symbolic instructions can introduce a text-action disconnect, frequently failing to activate the internal representations necessary for correct task execution. To address this, the paper introduces Neural Procedural Memory (NPM), a training-free framework that represents agent memory through implicit activation steering rather than explicit instructions. By distilling procedural skills from historical contrastive experiences into steering vectors in the activation space, NPM directly activates the task-relevant neural mechanisms to guide task execution. Evaluations across four agent benchmarks show that NPM performs comparably to baselines using explicit textual instructions. Furthermore, the results show that combining implicit steering with explicit workflows provides complementary advantages, leading to more robust task execution. Representational analyses indicate that these steering vectors encode consistent task logic, forming organized structures within the activation space. These findings suggest that implicit activation steering provides a promising approach for managing agent memory.