LLM 理论进展

Discovering stable neuron behavior across entire domains remains a challenge in mechanistic interpretability. Existing methods often rely on instance-level point estimates or computationally expensive procedures, which either obscure population-level variability or limit scalable domain-wide analysis. We present RACE (Residual Alignment for Consistency Estimation), a forward-pass statistical framework that evaluates the domain-wide functional consistency of Transformer neurons. Perturbation experiments demonstrate that RACE achieves superior domain specificity compared to gradient-based point estimates. Meanwhile, token-distribution-level results verify the association between the selected neurons and the target domain. Furthermore, its computational overhead is two orders of magnitude lower than that of gradient-based methods.

Effective human-AI interaction requires systems that dynamically adapt to a user's behavior and evolving understanding. When users interact with Large Language Models (LLMs), these models typically respond to prompts without sensing the user's immediate reactions. This lack of communicative synchrony can lead to information overload or leave confusion unresolved in real time. In this paper, we introduce Aura, a framework that enables LLM systems to dynamically modulate output based on a user's evolving emotions. Aura's Perception Module continuously estimates the user's emotional state from facial expressions. Our Policy Module then selects interventions through a probabilistic belief model. Finally, Aura's Generation Module uses parameter-efficient Low-Rank Adaptation (LoRA) adapters to produce contextually tailored responses mid-turn during response generation. We evaluated Aura in a within-subjects user study (N=20) on information-seeking tasks, where it achieved statistically significantly higher normalized perceived learning gains than a Llama-3 baseline and reduced interaction time by 21% relative to existing LLM baselines (GPT-4o, Llama-3). Our results indicate that real-time, context-sensitive interventions can improve learning efficiency and user satisfaction without observable degradation in factual accuracy. Aura thus supports the potential for more responsive and effective human-AI interaction.

Large language models are increasingly used for knowledge graph question answering (KGQA), but can fail to correctly ground answers in the underlying graph. Current approaches to LLM-based KGQA either rely on full semantic parsing into executable queries such as SPARQL, which is brittle in practice due to complex schemas or incompleteness of real-world KGs, or on LLM-reasoning and answer generation over KGs, which can be more robust but lacks formal guarantees. In this work, we study a complementary setting in which candidate answers are generated by an LLM-based system and subsequently verified using lightweight symbolic constraints derived from the question. We introduce Constrained Entity Selection under Partial Knowledge (CES-PK), a problem formulation that focuses on eliminating invalid answers and providing symbolic support for valid ones without requiring construction of executable logical forms. To account for incomplete KGs, we employ a three-valued constraint semantics (satisfied, violated, unknown) that avoids incorrect rejections under open-world assumptions. To demonstrate the effects of our method, we instantiate this framework over the Hetionet biomedical knowledge graph and evaluate the impact of type, relation, and exclusion constraints. Experiments show that precision improves by filtering invalid candidates, while recall is preserved due to retaining candidates whose constraints are not explicitly violated. Satisfied constraints provide additional positive symbolic evidence to rank remaining candidates.

Large Language Models (LLMs) have shown strong capabilities in table reasoning, but their effectiveness degrades as tables grow in size and complexity due to irrelevant context and difficulty localizing the evidence required for reasoning. Existing approaches typically reason over either the full table or a single reduced view, which can still obscure important row-column relationships. We introducePARTAB (Partition-Aware Reasoning overTables), a framework that constructs a structured evidence interface between the LLM and the table. PARTAB represents query-relevant evidence as semantically coherent, row-linked table regions and performs hierarchical selection over column groups and row-level partitions before composing the selected evidence for answer generation. We evaluate PARTAB on multiple table reasoning benchmarks, covering question answering, fact verification, and numerical reasoning. PARTAB consistently improves over full-table prompting and several recent table reasoning methods, achieving strong performance on WikiTableQuestions and TabFact while remaining competitive on numerical reasoning. Additional analyses show that semantic partitioning and targeted evidence selection improve evidence localization, substantially reduce the reasoning context, and provide larger benefits on complex tables. These results demonstrate the value of structured, partition aware evidence construction for scalable table reasoning.

Large language models (LLMs) are increasingly used for future prediction, motivating the use of multiple models as a wisdom-of-the-crowd mechanism. However, simply increasing crowd size does not guarantee effective diversity, as different LLMs may exhibit redundant behaviors. We propose a behavior-aware framework for constructing diverse LLM crowds. The framework characterizes models using their reasoning traces on independent development tasks, clusters models by behavioral similarity, and selects representatives for collective prediction. We evaluate 25 LLMs using seven development benchmarks for behavioral diversity modeling and two future-prediction benchmarks for evaluating diverse crowds' performance. Our results show that crowd composition can matter more than crowd size: a three-model medoid crowd based on K-means++ behavioral clustering outperforms conventional voting over all 25 models on both prediction benchmarks, while reducing model calls by 88% and inference cost by approximately 80%. The results further suggest that representative behavioral diversity, rather than simply maximizing diversity, is important for constructing effective LLM crowds

Systematic dialectal performance gaps in language models (LMs) are well documented, but the source of these disparities within the modern language modeling pipeline remains unclear. Our study traces this "dialect tax" across the natural language processing pipeline. Using parallel English dialect corpora that hold meaning fixed while varying surface form, we first confirm that LMs recognize matched Standard American English (SAE) and dialectal texts as semantically equivalent. However, we discover further representational gaps corresponding to downstream performance gaps. Across model families and generations, modern LMs still encode dialectal texts unequally during tokenization, pre-training, post-training, and inference. Strikingly, bypassing traditional subword segmentation via a character-level counterfactual tokenizer removes neither input and output asymmetries nor dialectal accuracy gaps. During pre-training, dialect pairs induce more divergent gradient updates than pairs of entirely unrelated SAE documents, indicating that models find semantically equivalent dialectal content harder to learn from than unrelated SAE documents. During post-training, reward models show contextual, unstable dialect preferences, assigning higher values to isolated AAVE-exclusive tokens than to SAE-exclusive tokens, while full reasoning contexts receive task- and model-dependent dialect penalties. Overall, our findings suggest that the dialect tax is encoded and accumulated not by any one step in isolation, but at every step of the language modeling process.

Grammatical knowledge and how it is empirically tested are typically considered robust to the frequency of the lexical items in the expressions. However, neural network-based models of grammaticality exhibit high sensitivity to lexical frequency. We draw upon Complementary Learning Systems theory to test the hypothesis that robustness to lexical frequency can arise via a hippocampal episodic memory mechanism, which enables rapid encoding and retrieval of specific experiences and allows learners to leverage them when processing rare patterns. We use retrieval-augmented language models as an instantiation of such an episodic memory mechanism (specifically, $k$-nearest-neighbor language models that augment parametric models with explicit instance storage), and test whether this augmentation helps close the lexical frequency gap that vanilla language models exhibit in syntactic contrast tests. Using syntactic contrasts with frequency-stratified test items, we find that retrieval augmentation narrows the performance gap between high- and low-frequency items, consistent with episodic memory compensating for weak parametric representations. This benefit is consistent across different syntactic phenomena and across models pretrained on child-realistic and large-scale data. Additionally, we show that structural information is critical for effective retrieval, whereas semantic similarity alone provides little benefit. While these are promising proof-of-concept results supporting our hypothesis, the frequency gap is narrowed rather than fully closed. Based on our analyses, we propose preferential reweighting of retrieved instances, better representations and retrieval strategies for structural information, and flexible configurations of storage and retrieval as promising future directions for improving the implementation of episodic memory in language models.

Aligning large language models to human-centered objectives is difficult when targets are non-executable and context-dependent, limiting reliable verification and scalable supervision. Although synthetic data expands coverage, weak verification shifts the bottleneck from generation to selection. Noisy signals destabilize iterative refinement and can cause silent regressions. We propose Agentic Data Evolution (ADE), a data-centric framework that organizes synthetic supervision as evolving data snapshots. ADE improves data snapshots through a closed-loop Observation-Variation-Selection (OVS) procedure, where a steady-state admission mechanism acts as a quality ratchet that conservatively gates updates for sustained cross-round improvement. We validate these improvements through complementary intrinsic trend tracking and extrinsic post-training evaluation. On DEV300, ADE raises the intrinsic win rate from 50% to 75.81% and the extrinsic win rate from 55.20% to 68.86%, consistent performance gains across diverse benchmarks. Blind expert evaluation further confirms this, with a 66.11% preference for evolved answers. These gains extend across post-training methods, model scales, and tasks beyond the target weakly verifiable educational objectives. Resources are available at https://github.com/ZeroLoss-Lab/Agentic-Data-Evolution.

Transformer-based language models are widely used as models of human language processing, yet their attention mechanisms allow lossless access to the full preceding context, unlike the limited memory systems of humans. We hypothesize that installing memory constraints into transformers' attention mechanisms can improve their fit to human behavioral data. While previous work has explored individual constraints in isolation, we conduct a systematic comparison of multiple attention-based memory mechanisms across different model sizes and training corpora, evaluating both psychometric predictive power for human reading times and grammatical competence. We additionally compare static constraints, in which the constraint strength is fixed throughout training, to dynamic memory curricula. We find that constraints that are sensitive to the content of intervening tokens consistently achieve the highest alignment with human reading times, outperforming distance-based constraints. We observe a dissociation between psychometric fit and grammatical competence under dynamic memory curricula, suggesting that Transformers cannot serve as a one-size-fits-all cognitive model.

Data mixing is a central design problem in large language model pretraining: given a fixed token budget, practitioners must decide how much data to allocate to each domain. Recent proxy-based methods address this problem by training small models on candidate mixtures, fitting a response model, and using the response to select mixtures for larger-scale training. We show that this workflow has the structure of a classical mixture experiment. Under this view, data domains are mixture components, token shares are component proportions, proxy-training runs are experimental design points, and validation loss defines a response surface over the probability simplex. We develop this formulation using sparse second-order Scheffé response-surface models and construct model-robust $\mathcal{I}$-optimal designs for proxy data-mixing experiments. Using RegMix as an empirical case study, we demonstrate how the framework can both interpret observed mixture responses and design more efficient proxy experiments. The Scheffé analysis shows that domain value is strongly relational: several domains that are weak under additive effects become favourable through pairwise interactions, especially through combinations with web-derived text. The sparse Scheffé model preserves mixture rankings across model scales and remains competitive with a flexible machine-learning predictor while providing an explicit decomposition of additive and interaction effects. In a simulation study calibrated to observed proxy-training responses, model-robust $\mathcal{I}$-optimal designs recover the relevant mixture ordering after removing about 25% of the original proxy runs. These results suggest that LLM data mixing should be treated not only as a prediction problem, but also as an experimental-design problem in which the proxy mixtures themselves can be chosen to improve statistical efficiency.

Large Language Models (LLMs) demonstrate strong capabilities in automated essay scoring (AES), but contemporary approaches typically employ fixed prompt selection, failing to address operational cost concerns and evolving optimal configurations. We propose a novel cost-aware approach that treats each prompt type as an arm in a multi-armed bandit (MAB) controller, enabling adaptive selection of optimal prompting strategies during inference. Our experiments on IELTS Writing Task 2 essays show that the MAB framework achieves comparable scoring accuracy to exhaustive grid search while reducing LLM calls by 78.4% to find the best grading approach. We implemented four distinct grading recipes (multi-step vs. single-step assessment, with vs. without calibration examples) and found that the multi-step approach with examples achieves the highest accuracy. By tracking token usage and latency alongside agreement metrics, we produce the first cost-reliability learning curves for essay scoring, providing actionable insights for educational technology platforms that must balance operational costs against assessment validity. This work represents the first application of online control mechanisms to adaptively select prompting strategies in AES, transforming prompt selection from an offline hyperparameter optimization problem into an efficient online learning task.

Large language models (LLMs) are known to exhibit social sycophancy, often validating or agreeing with users in socially sensitive contexts. Existing evaluations typically measure sycophancy under a fixed prompt formulation, leaving unclear whether such behavior is stable when the same underlying situation is presented with different sycophancy-relevant prompt variants. In this work, we study sycophancy prompt sensitivity: the extent to which changes in user confidence, emotional framing, social consensus, or validation-seeking language alter a model's sycophantic behavior. We refer to our evaluation framework as SyPS, short for Sycophancy Prompt Sensitivity. Building on existing social sycophancy evaluation settings, SyPS constructs controlled prompt variants that preserve the same underlying user situation while varying sycophancy-relevant social cues. We introduce the Sycophancy Prompt Sensitivity Score (SPSS), an instance-level measure of sycophancy variation across paired prompt variants. Unlike aggregate sycophancy rates, SPSS separates baseline sycophancy from prompt-induced shifts, enabling model-level comparisons of robustness to sycophancy-relevant social cues. Empirically, we find that sycophancy prompt sensitivity is socially structured: validation-seeking and emotional-pressure cues often increase sycophancy, whereas counter-framing and anti-sycophancy prompts tend to reduce it. Our framework highlights whether LLMs maintain stable social judgments while adapting appropriately in tone.

Large language models (LLMs) are becoming routine instruments of scientific research, assisting with literature synthesis, hypothesis development, coding, and formal reasoning. Their use raises a central epistemic question: when parts of scientific reasoning are delegated to an artificial system, what conditions must remain under human control for the resulting knowledge claims to retain epistemic legitimacy and accountable authorship? This paper develops a normative and conceptual framework for analyzing such delegation. Scientific reasoning is treated as a distributed process in which the origin of a contribution may vary between human and machine, while responsibility for its acceptance into the scientific record remains human. The framework distinguishes content origin $O(g)$, completion of human verification $V(g)$, responsibility assignment $R(g)$, accountable human ownership $M(g)$, and epistemic outcome $E(g)$. These constructs separate the provenance of a claim from the process by which it is checked, the epistemic outcome of that checking, and the human responsibility attached to its disposition. The central proposition is that the ethical boundary of LLM-assisted research is determined primarily by adequate verification and accountable human ownership rather than by the degree of machine involvement itself. On this basis, the paper develops the notion of an epistemic audit: a structured record of delegation, verification, provenance, and responsibility intended to make AI-assisted reasoning transparent and reviewable. The resulting framework provides a formal vocabulary for distinguishing responsible cognitive delegation from the transfer or neglect of epistemic responsibility in scientific research.

Large Language Models (LLMs) recognise patterns but do not natively track the path of exclusions that a coherent discourse demands. When an input rests on a fabricated authority, a misapplied mechanism, or a surreptitious analogy, an unsteered LLM tends to engage with it as if it were grounded, and to reify the misstep into any structured output it generates. Logic-Augmented Generation (LAG) with POLANYI++, an LLM-steering method that uses heuristics, ontologies and problem solving methods for tacit knowledge extraction, produces an Extended Knowledge Graph (XKG) in OWL2, but inherits the same vulnerability: a sophisticated nonsensical input is reified into the graph alongside the legitimate triples, and is hardly detectable by automated reasoners since the XKG is generated jointly with the wrong assumptions. We introduce DARKSIDE, a coherence auditing method on top of POLANYI++. It formalises the trail as an explicit data structure of accumulated exclusions over discourse time, complemented by a warrant axis that classifies each named referent as Warranted, Unattested, Misattributed or Fabricated, with an escalation rule that pushes the DelegationRiskAssessment to UNSAFE when the fabricated rate is positive or the unsupported rate exceeds a threshold. We evaluate DARKSIDE as a steering layer over a Gemini 3 on BSBench, a 100-item adversarial corpus of sophisticated-sounding nonsense across software engineering, finance, healthcare, physics and law, with Claude Sonnet 4.6 as an independent judge. The empirical evidence supports an architectural claim: when an LLM forward pass is wrapped in an ontology-mediated negative-trail apparatus, the structural pattern-vs-path gap can be partially scaffolded. The XKG functions as the missing memory, and the warrant axis as an epistemic firewall.

Arabic Natural Language Processing (NLP) has grown rapidly over the past decade, driven by digital transformation in the Arab world, social media, and large language models (LLMs). Despite this growth, a comprehensive quantitative meta-analysis remains absent. This study presents a bibliometric and topic-based analysis of 7,120 Arabic NLP papers published between 1960 and 2026, sourced from five platforms (arXiv, ACL Anthology, Semantic Scholar, Crossref, OpenAlex) plus an additional targeted OpenAlex subset. We employ BERTopic for topic modeling, regression analysis, social network analysis, and geographic mapping. Our findings show a significant publication surge after 2020, driven by transformer models and LLMs. Topic modeling identifies 19 themes, the largest centered on text, speech, translation, and recognition (2,942 papers). Citation analysis reveals a positive correlation between paper age and citations (r = 0.245, p < 0.001); regression (R^2 = 0.105) shows that indexing in OpenAlex or Semantic Scholar and institutional affiliation are associated with higher citations. Saudi Arabia, the United States, and Egypt lead in research output. A task-dialect gap matrix identifies understudied areas, including summarization for Maghrebi, Iraqi, and Sudanese dialects. The largest topic has the highest H-index (90), followed by sentiment analysis (57). Our quantitative approach complements existing qualitative surveys and offers recommendations to prioritize under-resourced dialects and develop culturally aligned benchmarks.

Large language models (LLMs) require effective unlearning to address privacy regulations and safety concerns. However, achieving precise forgetting without compromising general utility remains challenging. Existing sequence- and token-level methods penalize target outputs without modeling their context-dependent retrieval paths, which can disrupt linguistic structure or suppress benign knowledge. We present ADU, a fine-grained, training-based framework that shifts unlearning from token erasure to contextual attention-pathway decoupling. Exploiting the functional distinction between local and global attention heads, ADU identifies preplan positions that retrieve persistent sensitive anchors and fixes their candidate paths under the original model. It then trains attention-projection adapters to suppress attention mass along these paths while preserving local-attention structure and retain-set language modeling. Post-training activation exchange tests whether the modified attention-output module transmits the learned forgetting effect. ADU achieves the strongest aggregate performance among evaluated baselines on the TOFU and WMDP benchmarks, including a Forget Quality of (0.93) on TOFU. It preserves 87--98% of model utility (92.9% on average versus 81.9% for baselines) while reducing side effects in benign contexts.

The alignment of Large Language Models heavily relies on English-centric high-quality preference data, which often leads to suboptimal performance in other languages. In this paper, we propose Cross-lingual Ranking Preference Optimization~(CRPO), a novel framework that leverages robust preference knowledge from English to facilitate preference alignment in the target language. We design a hierarchical structure within parallel preference pairs across the target language and English to jointly optimize intra- and inter-lingual preferences, thereby enhancing language adaptation and output quality. Building on the LambdaLoss framework, CRPO goes beyond the binary comparison based optimization by providing a relative ranking signal across multiple candidate responses. Our experiments across five languages with varying resource scales demonstrate that CRPO consistently outperforms standard approaches in both instruction-following and knowledge utilization capability. Notably, the robust performance gains observed across various weighting schemes further validate the empirical effectiveness of our hierarchical design in a multilingual setup. Furthermore, our findings highlight that CRPO significantly improves both reward margins and the log-probability of desirable responses, contributing to a more stable preference manifold for cross-lingual alignment. Our code is available at https://github.com/dltmddbs100/CRPO.

Large language models (LLMs) are increasingly used in investment decision-making, yet prior work shows that they exhibit systematic, model-specific investment preferences. We study whether a model's overall investment stance can be calibrated to a specified direction and strength. We introduce an investment-bias dial, an inference-time intervention on a single neuron that continuously adjusts a model-level decision prior---its overall tendency toward buying or selling---without targeting specific firms or investment attributes. Using matched positive and negative evidence, we evaluate five open-weight LLMs and find that the dial produces monotonic changes in investment stance without modifying prompts or model parameters. At the response level, the dial shifts both investment decisions and the evidential emphasis of generated rationales under identical inputs. In an agentic retrieval setting, the dial also changes what information the model searches for, which evidence it selects, and which evidence is reflected in its final analysis. In a long-context evaluation, the dial maintains stable stance control as context length increases, whereas a matched system-prompt instruction progressively attenuates. We further show that changes in the dial propagate to security rankings and downstream portfolio composition in an exploratory backtest. Overall, our results show that an LLM's aggregate investment stance can be calibrated toward a specified target at inference time.

Spoken Language Models (SLMs) generate textual responses directly from speech, offering an alternative to cascaded systems. Despite recent advances, existing SLMs still exhibit weaker instruction-following behavior and limited generalization across diverse tasks compared to text-based language models. Our analysis shows that speech and text representations in current SLMs remain weakly aligned despite strong downstream performance, indicating that structural differences between continuous, temporally varying speech and discrete text remain insufficiently addressed. To address this, we propose a simple framework that decouples length mismatch from semantic alignment and encourages closer correspondence between speech and text representations. Experiments across multiple benchmarks demonstrate competitive performance against strong baselines, underscoring the importance of explicitly addressing structural differences between speech and text in SLM training. Our code is publicly available at https://github.com/jaykim9870/Do_SLMs_Hear_Speech_as_They_Read_Text.

In this technical report, we present a training-free framework for audio-guided video object segmentation, which integrates Multimodal Large Language Models (MLLMs) with SAM-based segmentation models. We decompose the task into several stages and identify suitable foundation models for each stage. Without introducing additional model training or task-specific fine-tuning, our approach leverages the strong multimodal reasoning capabilities of MLLMs to model text-visual correspondence and employs SAM-based models for accurate object mask generation. The proposed framework demonstrates the effectiveness of leveraging foundation models for audio-guided video segmentation and achieves competitive performance in the MeViS-Audio Track of the 8th LSVOS Challenge.

Mixed-integer programming (MIP) lies at the core of operations research and industrial optimization. While large language models (LLMs) have recently shown promise in automated MIP modeling from natural language, they prioritize semantic correctness but overlook formulation strength, severely bottlenecking the efficiency of downstream solvers. We propose FormuEvo, an LLM-guided evolutionary framework for automated discovery of solver-efficient MIP formulations. FormuEvo frames MIP formulation design as evolutionary optimization over the symbolic space of MIP formulations, represented as executable modeling programs, by iteratively generating, evaluating, and selecting stronger candidates via LLM-driven crossover, mutation, and repair operations. To move beyond blind exploration, FormuEvo introduces a solver-informed diagnosis mechanism that exploits fine-grained solver statistics as verbal gradients for targeted refinement. Additionally, a structured memory abstracts prior experience into reusable modeling strategies, avoiding redundant exploration while enabling zero-shot transfer to unseen problems and bootstrapping smaller LLMs. Experiments across diverse linear and non-linear problems demonstrate that FormuEvo discovers formulations that significantly outperform both expert-designed formulations and existing LLM-based approaches, accelerating solvers by up to 5.5$\times$, with distilled knowledge transferring effectively across problems and model scales.

Large Language Models generate text autoregressively, but open-ended generation is prone to verbatim looping, in which models repeat spans already present in context. Standard defenses such as repetition, presence, and frequency penalties and n-gram blocking act on token recurrence rather than the sequential structure of a loop, and often suppress looping only at strengths that also degrade formatting or fluency. We propose Don't Repeat Yourself (DRY), a sampling-time logit adjustment that penalizes a candidate token only when generating it would extend the current suffix into an exact continuation of a span seen earlier in the context. Sequence breakers protect chat templates and formatting tokens. Across models from 1.5B to 120B parameters, nine prompt families, and a 600-pair human study, DRY reduces suffix-extension rate by 47% while improving lexical diversity. An intervention-matched placebo produces no comparable reduction, identifying suffix matching as the operative mechanism. On AWQ-quantized 70B and 120B models, DRY reduces loop rate by roughly half while preserving MT-Bench, MMLU, and GSM8K performance, whereas standard alternatives lose measurable ground. DRY has been adopted by popular open-source LLM inference frameworks including llama.cpp, ExLlamaV2, and text-generation-webui, highlighting its practical impact on text generation.

Composed Image Retrieval (CIR) is an emerging paradigm in content-based image retrieval that enables users to formulate compositional queries by combining a reference image with an auxiliary modality, usually text-based. This approach supports fine-grained search where the target image shares structural elements with the user-provided image while incorporating the modifications specified by the auxiliary text. Conventional CIR methods rely on multimodal fusion to combine visual and textual features into a joint query embedding, which requires training modules that align composed queries with the targets. In this work, we propose PeFuse (for pseudo-fusion), a training-free framework that leverages pretrained Diffusion Models and Multimodal Large Language Models to bridge modalities via generative conversion. We introduce two novel strategies: uni-directional and bi-directional conversion, which convert CIR into four single-modality retrieval problems. These methods reformulate CIR as either intra-modal or cross-modal single-query retrieval tasks, bypassing the need for dedicated task-specific training. Extensive experiments on standard benchmarks demonstrate that converting CIR into text-to-image retrieval tasks is more effective than alternative conversion strategies, achieving competitive or superior performance compared with state-of-the-art methods, while maintaining high flexibility thanks to replaceable components of the conversion pipeline. These results highlight the effectiveness of the pseudo-fusion paradigm for zero-shot CIR. Our code is publicly available at: https://github.com/StevenXuf/PeFuse4CIR.

Recent work proposes next-chunk reasoning RL for leveraging no-CoT data---corpora such as worked solutions and textbook derivations that contain reasoning-rich content but lack explicit chain-of-thought annotations. The method trains a model to generate implicit reasoning traces and rewards them by their ability to predict the next chunk of text. While promising, existing evaluations primarily compare against conventional SFT baselines, leaving open whether the gains come from the RL formulation itself or from more effectively exposing the model to no-CoT data. We address this question with a controlled study of next-chunk reasoning RL and a simple but previously overlooked alternative: Mixed SFT, a single supervised fine-tuning stage that jointly trains on no-CoT and long-CoT data. Despite its simplicity, Mixed SFT achieves a clearly higher post-RLVR performance ceiling than next-chunk reasoning RL while requiring over 60 times less training compute. The advantage is consistent across in-domain mathematical reasoning and out-of-domain reasoning tasks. Moreover, we show that higher pre-RLVR accuracy does not necessarily translate into higher post-RLVR accuracy, highlighting the need to evaluate no-CoT training strategies in the context of the full post-training pipeline.

Digital Twin (DT) ecosystems integrate heterogeneous computational models to represent complex systems under evolving, purpose-specific objectives. Systematic reuse of existing high-quality models and datasets is essential for scalable DT development, yet is constrained by heterogeneity in semantic intent, data structures, behavioral interfaces, and execution environments. As a result, integration becomes a cross-model, cross-view consistency problem that is hard to predict, quantify, and compare across design choices. Existing standards and integration platforms address these concerns separately, offering limited support for structured, purpose-aware compatibility assessment and early feasibility analysis when models are reused under new DT objectives. This paper introduces a multi-viewpoint integration modeling framework grounded in the Reference Model of Open Distributed Processing (RM-ODP). The framework structures integration-relevant knowledge across domain, information, computational, engineering, and technology viewpoints, representing cross-view dependencies as explicit, machine-actionable metadata. It comprises (i) a viewpoint-structured Model Metamodel for systematic model description and discovery, and (ii) a pattern-aware Mismatch Detector that operationalizes cross-view compatibility constraints via integration patterns, combining deterministic rule generation with Large Language Model (LLM)-assisted reasoning. This enables systematic identification of semantic, informational, and runtime inconsistencies and supports reasoning about integration feasibility and effort before implementation. Expert validation and an environmental modeling case study show that the approach enables structured compatibility reasoning, improves transparency of integration assumptions, strengthens cross-view interoperability, and supports scalable reuse in heterogeneous DT ecosystems.

Spatial multi-omics technologies jointly profile gene expression, surface proteins, and histology at each tissue spot, yet most spatial domain discovery methods provide only cluster assignments, without indicating assignment reliability, modality contributions, or why a domain decision should be trusted. We present OmicSync, a reliability-aware spatial multi-omics framework that couples unsupervised domain clustering with evidence-constrained LLM reasoning using model-derived per-spot signals, including assignment confidence, epistemic routing uncertainty, and modality-routing weights. These signals are converted into structured evidence dictionaries and used to generate standard, stepwise, counterfactual, contrastive, and uncertainty-focused explanations. OmicSync integrates a KAN-GCN backbone with spatial encoding, cross-modal fusion, uncertainty-aware routing, cell-type supervision, and missing-modality imputation. We further introduce OmicSync-R, which closes the reasoning-clustering loop by using automatically computed reasoning-quality scores as REINFORCE rewards, allowing reasoning coherence to shape the latent structure without backpropagating through the language model. Across four 10x CytAssist FFPE spatial proteomics benchmarks, OmicSync achieves the best average rank on Human Tonsil (1.44), Glioblastoma (1.78), and Tonsil Add-on (1.22), and second-best on Human Breast Cancer (2.33). OmicSync-R further improves ARI on Human Breast Cancer from 45.73 to 46.72 and outperforms existing methods on six of nine clustering metrics. Together, OmicSync and OmicSync-R enable reliability-aware, spot-level auditable spatial domain discovery guided by evidence-constrained reasoning.

Although Large Language Models (LLMs) are aligned to optimize for both helpfulness and harmlessness, these dual objectives may conflict, inevitably leading to alignment failures. This work systematically investigates instances where LLMs fail to exhibit ethical behavior. To understand the underlying mechanics of these vulnerabilities, we introduce a probing methodology that presents unethical scenarios to LLMs in three distinct structural modalities: objective classification tasks, subjective first-person statements, and direct requests for assistance. We find that model performance degrades in the request-for-assistance-based form. Using Layer-wise Relevance Propagation (LRP), we trace this discrepancy to an attribution bias: the model places greater emphasis on benign task-framing tokens (e.g., "Can you help me...") than on tokens signaling the underlying unethical behavior (e.g., "without getting caught"), which we term cue-tokens. We hypothesize that this under-attribution contributes to harmful compliance. To test this, we introduce two LRP-guided decoding methods that steer generation toward trajectories more relevant to cue tokens. Empirical evaluations show that these interventions promote safer responses, supporting cue-token attribution's role in compliance failures.

Multimodal Large Language Models (MLLMs) have demonstrated great performance, yet they often suffer from severe modality misalignment when confronted with densely constrained spaces for interior design. Due to the loss of high-frequency local topological details and fine-grained aesthetic shifts during visual encoding, existing MLLMs frequently hallucinate, yielding physical spatial collisions and visual aesthetic dissonance. To address this, we propose Dual-prior Activation Residual Task-vectors Injection mechanism (DART-I) for MLLMs. It shifts the paradigm from lossy text-prompting to direct latent intervention, utilizing weak-to-strong deterministic rules to anchor the causal reasoning of MLLMs for interior design. Specifically, DART-I operates in three steps: it first explicitly extracts continuous spatial distance and color typography features from images using extremely lightweight weak experts; subsequently, it transforms these deterministic priors into directional task vectors via a linear projection network; these vectors are dynamically injected as residual terms into the latent space of the frozen MLLMs, steering MLLMs towards precise reasoning for interior design. Stepping outside the conventional paradigms, our method achieves precise reasoning without fine-tuning the MLLMs, effectively bypassing expensive computational costs and catastrophic forgetting. Extensive experiments on various benchmarks demonstrate the effectiveness and advantages of DART-I.

Large language model (LLM) agents rely heavily on knowledge encoded in model parameters or presented as unstructured context. In domain-specific tasks, this leaves important semantic connections implicit. This often results in incomplete evidence use and brittle multi-step decisions. Ontologies offer a way to externalize domain concepts and relations as machine-interpretable structures, but constructing task-usable ontologies traditionally requires substantial effort from domain experts and is difficult to scale. Automatic construction is also challenging: an ontology that appears semantically plausible may not contain the relational structures needed for actual decision making. We present OaK, an ontology-as-a-kernel framework that dynamically constructs and refines task-oriented ontologies for LLM agents. Given task requirements and training data, OaK constructs an ontology and its knowledge graph, generates task-adaptation functions for graph reasoning, and uses judge feedback to iteratively refine both. By making relevant concepts and relations explicit, the ontology grounds knowledge retrieval and multi-step decision making. We evaluate OaK on TravelPlanner, CRMArenaPro, and ToolQA. Results show that OaK improves standard LLM agents, strengthens evidence grounding, and boosts the reliability of multi-step reasoning.

Large language models increasingly operate over large collections of tools, functions, APIs, and specialized agents. As the candidate action space grows, a function-calling model must process more schemas, consume more prompt tokens, and distinguish among increasingly similar or irrelevant alternatives. We study a complementary systems strategy: reduce the candidate set before language-model inference while leaving the downstream model unchanged. We introduce AgentWeave, a deterministic pre-inference routing layer that constructs a bounded model-visible action space using eligibility, requirement, capability, and routing signals. We evaluate AgentWeave with a frozen BFCL-derived routing-pressure protocol using the public MadeAgents/Hammer2.1-1.5b model. On 48 fresh BFCL V4 multiple-function tasks, AgentWeave achieves 6/48 (12.5%) native BFCL successes, whereas all-tools, deterministic random top-8, and semantic top-8 baselines each achieve 0/48. The paired success difference is +12.5 percentage points with a 10,000-resample paired bootstrap 95% confidence interval of +4.17 to +22.92 points and exact McNemar p=0.03125. Relative to all-tools exposure, AgentWeave presents 70.18% fewer tools, uses 61.70% fewer input tokens, and exhibits 50.95% lower mean local-model latency. The result is deliberately narrow: this is a BFCL-derived routing-pressure study rather than an official full BFCL leaderboard score, and absolute task success remains low. The evidence nevertheless shows that candidate-space construction can materially affect a fixed model's function-calling behavior and motivates evaluating routing as a distinct stage before model reasoning.