Papers

Large language models increasingly mediate multilingual professional communication, where useful generation requires adapting to community conventions about which expressions are retained, translated, or mixed. Existing benchmarks rarely isolate such community-conditioned choices. We introduce ChiEngMixBench, a controlled benchmark for Chinese AI/CS discourse, where Chinese frames routinely incorporate established English technical terms. Built from public technical discussions, it contains 1,706 source-derived candidate pairs covering 1,344 non-empty normalized terms, including a 1,167-pair strict subset that fixes the Chinese prefix and syntactic position while varying only the terminology form. The benchmark combines paired likelihood comparisons with a transparent reference-profile diagnostic for open-ended responses. Across nine open-weight models, Chinese equivalents receive higher likelihood on most pairs, revealing a gap between source-attested usage and model preference. Specialized terms show a small directional lift that is not robust after frequency and length controls and multiple-comparison correction. Human evaluation and baseline analyses show that reference-profile conformity is informative under the intended mixed-style rubric but does not reliably predict holistic response preference. ChiEngMixBench provides a reusable testbed for community-specific multilingual conventions with explicit diagnostic boundaries.

Large language models achieve strong reasoning performance, yet existing decoding strategies either explore blindly (random sampling) or redundantly (independent multi-sampling). We propose Entropy-Tree, a tree-based decoding method that exploits entropy as a signal for branching decisions–expanding the search tree only at positions where the model exhibits genuine uncertainty. Entropy-Tree shows superior accuracy and calibration in reasoning tasks: it achieves better pass@k than Multi-chain across multiple models and datasets, and its predictive entropy demonstrates better AUROC compared to several traditional metrics. Entropy-Tree unifies efficient structured exploration and reliable uncertainty estimation within a single decoding procedure.

Large Language Models (LLMs) have demonstrated remarkable abilities to solve problems requiring multiple reasoning steps, yet the internal mechanisms enabling such capabilities remain elusive. Unlike existing surveys that primarily focus on engineering methods to enhance performance, this survey provides a comprehensive overview of the mechanisms underlying LLM multi-step reasoning. We organize the survey around a conceptual framework comprising seven interconnected research questions, from how LLMs execute implicit multi-hop reasoning within hidden activations to how verbalized explicit reasoning remodels the internal computation. Finally, we highlight five research directions for future mechanistic studies.

Large language models (LLMs) have been widely used for mental health support. However, current safety evaluations in this field are mostly limited to detecting whether LLMs output prohibited words in single-turn conversations, neglecting the gradual erosion of safety boundaries in long dialogues. Examples include making definitive guarantees, assuming responsibility, and playing professional roles. We believe that with the evolution of mainstream LLMs, words with obvious safety risks are easily filtered by their underlying systems, while the real danger lies in the gradual transgression of boundaries during multi-turn interactions, driven by the LLM's attempts at comfort and empathy. This paper proposes a multi-turn stress testing framework and conducts long-dialogue safety tests on three cutting-edge LLMs using two pressure methods: static progression and adaptive probing. We generated 50 virtual patient profiles and stress-tested each model through up to 20 rounds of virtual psychiatric dialogues. The experimental results show that violations are common, and both pressure modes produced similar violation rates. However, adaptive probing significantly advanced the time at which models crossed boundaries, reducing the average number of turns from 9.21 in static progression to 4.64. Under both mechanisms, making definitive or zero-risk promises was the primary way in which boundaries were breached. These findings suggest that the robustness of LLM safety boundaries cannot be inferred solely through single-turn tests; it is necessary to fully consider the wear and tear on safety boundaries caused by different interaction pressures and characteristics in extended dialogues.

We present a theory-first framework that interprets inference-time adaptation in large language models (LLMs) as online Bayesian state estimation. Rather than modeling rapid adaptation as implicit optimization or meta-learning, we formulate task- and context-specific learning as the sequential inference of a low-dimensional latent adaptation state governed by a linearized state-space model. Under Gaussian assumptions, adaptation follows a Kalman recursion with closed-form updates for both the posterior mean and covariance. This perspective elevates epistemic uncertainty to an explicit dynamical variable. We show that inference-time learning is driven by covariance collapse, i.e., rapid contraction of posterior uncertainty induced by informative tokens, which typically precedes convergence of the posterior mean. Using observability conditions on token-level Jacobians, we establish stability of the Bayesian filter, prove exponential covariance contraction rates, and derive mean-square error bounds. Gradient descent, natural-gradient methods, and meta-learning updates arise as singular, noise-free limits of the filtering dynamics, positioning optimization-based adaptation as a degenerate approximation of Bayesian inference. The resulting theory provides a unified probabilistic account of in-context learning, parameter-efficient adaptation, and test-time learning without parameter updates. It yields explicit guarantees on stability and sample efficiency, offers a principled interpretation of prompt informativeness via information accumulation, and clarifies the role of uncertainty dynamics absent from existing accounts. Minimal illustrative experiments corroborate the qualitative predictions of the theory.

The increasing prevalence of sparse Mixture-of-Experts (MoE) architectures in large language models raises important questions regarding their reliability under stochastic decoding. While conditional computation enables substantial gains in computational efficiency, it remains unclear whether the interaction between sparse routing and temperature-based sampling compromises output stability relative to dense architectures. This work investigates whether conditional computation in MoE models amplifies decoding-induced randomness, leading to reduced reliability as temperature increases. We evaluate three representative models: OLMoE-7B (sparse base), Mixtral-8x7B (sparse instruction-tuned), and Qwen2.5-3B (dense instruction-tuned) on deterministic arithmetic reasoning tasks with objectively verifiable answers. Experiments span four decoding configurations, ranging from greedy decoding to T=1.0. Our evaluation encompasses accuracy, format compliance, output consistency across repeated generations, and confidence metrics, totaling 9,360 model generations. Results demonstrate that the sparse instruction-tuned model exhibits stability comparable to the dense instruction-tuned model across all decoding temperatures, while the sparse base model shows systematic degradation as temperature increases. These findings indicate that instruction tuning, rather than architectural sparsity, is the primary determinant of robustness to decoding randomness on deterministic tasks. We discuss the implications of these results for deploying sparse language models in reliability-critical applications, highlighting scenarios in which sparse architectures can be safely adopted without sacrificing output stability.

Bounded-context agents fail when intermediate reasoning exceeds an effective working-memory budget. We study compressed query delegation (CQD): (i) compress a high-dimensional latent reasoning state into a low-rank tensor query, (ii) delegate the minimal query to an external oracle, and (iii) update the latent state via Riemannian optimization on fixed-rank manifolds. We give a math-first formulation: CQD is a constrained stochastic program with a query-budget functional and an oracle modeled as a noisy operator. We connect CQD to classical rate-distortion and information bottleneck principles, showing that spectral hard-thresholding is optimal for a natural constrained quadratic distortion problem, and we derive convergence guarantees for Riemannian stochastic approximation under bounded oracle noise and smoothness assumptions. Empirically, we report (A) a 2,500-item bounded-context reasoning suite (BBH-derived tasks plus curated paradox instances) comparing CQD against chain-of-thought baselines under fixed compute and context; and (B) a human "cognitive mirror" benchmark (N=200) measuring epistemic gain and semantic drift across modern oracles.

Understanding affective polarization in online discourse is crucial for evaluating the societal impact of social media interactions. This study presents a novel framework that leverages large language models (LLMs) and domain-informed heuristics to systematically analyze and quantify affective polarization in discussions on divisive topics such as climate change and gun control. Unlike most prior approaches that relied on sentiment analysis or predefined classifiers, our method integrates LLMs to extract stance, affective tone, and agreement patterns from large-scale social media discussions. We then apply a rule-based scoring system capable of quantifying affective polarization even in small conversations consisting of single interactions, based on stance alignment, emotional content, and interaction dynamics. Our analysis reveals distinct polarization patterns that are event dependent: (i) anticipation-driven polarization, where extreme polarization escalates before well-publicized events, and (ii) reactive polarization, where intense affective polarization spikes immediately after sudden, high-impact events. By combining AI-driven content annotation with domain-informed scoring, our framework offers a scalable and interpretable approach to measuring affective polarization. The source code is publicly available at: https://github.com/hasanjawad001/llm-social-media-polarization.

Verifying whether a language model is genuinely reasoning or pattern-matching remains an open problem: learned verifiers are expensive, and output-based heuristics are brittle. We show that valid mathematical reasoning induces a measurable, training-free spectral signature in transformer attention. By treating each attention matrix as a weighted token graph, we extract four diagnostics: Fiedler value, High-Frequency Energy Ratio (HFER), spectral entropy, and smoothness, that require no learned parameters. Experiments across seven models from four architectural families yield effect sizes up to Cohen's $d = 3.30$ ($p < 10^{-116}$), enabling $85$–$96\%$ single-threshold classification accuracy. Two findings sharpen the interpretation. First, Platonic validity: the spectral signal tracks logical coherence rather than compiler acceptance, proofs rejected for timeouts or missing imports are correctly classified as valid, a distinction confirmed by a manual audit ($\kappa = 0.82$, $n = 51$). Second, architectural determinism: Sliding Window Attention shifts the discriminative feature from HFER to smoothness ($d = 2.09$, $p < 10^{-48}$), showing that attention design governs which spectral channel encodes reasoning quality. Causal ablation confirms the signature traces induction-head circuits. The method generalises to informal chain-of-thought ($d = 0.78$, $p < 10^{-3}$), and in proof search, HFER reranking improves Best-of-16 Pass@1 by $+4.4$–$6.6$%, matching $98\%$ of the AUC of fully supervised probes with zero labels. Spectral graph analysis is a principled, architecture-aware primitive for reasoning verification.

Population-based cancer registries depend on pathology reports as their primary diagnostic source, yet manual abstraction is resource-intensive and contributes to delays in cancer data. While transformer-based NLP systems have improved registry workflows, their ability to generalize across jurisdictions with differing reporting conventions remains poorly understood. We present the first cross-provincial evaluation of adapting BCCRTron, a domain-adapted transformer model developed at the British Columbia Cancer Registry, alongside GatorTron, a biomedical transformer model, for cancer surveillance in Canada. Our training dataset consisted of approximately 104,000 and 22,000 de-identified pathology reports from the Newfoundland & Labrador Cancer Registry (NLCR) for Tier 1 (cancer vs. non-cancer) and Tier 2 (reportable vs. non-reportable) tasks, respectively. Both models were fine-tuned using complementary synoptic and diagnosis focused report section input pipelines. Across NLCR test sets, the adapted models maintained high performance, demonstrating transformers pretrained in one jurisdiction can be localized to another with modest fine-tuning. To improve sensitivity, we combined the two models using a conservative OR-ensemble achieving a Tier 1 recall of 0.99 and reduced missed cancers to 24, compared with 48 and 54 for the standalone models. For Tier 2, the ensemble achieved 0.99 recall and reduced missed reportable cancers to 33, compared with 54 and 46 for the individual models. These findings demonstrate that an ensemble combining complementary text representations substantially reduce missed cancers and improve error coverage in cancer-registry NLP. We implement a privacy-preserving workflow in which only model weights are shared between provinces, supporting interoperable NLP infrastructure and a future pan-Canadian foundation model for cancer pathology and registry workflows.

Large Language Models (LLMs) have become a mainstay for many everyday applications. However, as data evolve their knowledge quickly becomes outdated. Continual learning aims to update LLMs with new information without erasing previously acquired knowledge. Although methods such as full fine-tuning can incorporate new data, they are computationally expensive and prone to catastrophic forgetting, where prior knowledge is overwritten. Memory-augmented approaches address this by equipping LLMs with a memory bank, that is an external memory module which stores information for future use. However, these methods face a critical limitation, in particular, the memory bank constantly grows in the real-world scenario when large-scale data streams arrive. In this paper, we propose MBC, a model that compresses the memory bank through a codebook optimization strategy during online adaptation learning. To ensure stable learning, we also introduce an online resetting mechanism that prevents codebook collapse. In addition, we employ Key-Value Low-Rank Adaptation in the attention layers of the LLM, enabling efficient utilization of the compressed memory representations. Experiments with benchmark question-answering datasets demonstrate that MBC reduces the memory bank size to 0.3% when compared against the most competitive baseline, while maintaining high retention accuracy during online adaptation learning. Our code is publicly available at https://github.com/Thomkat/MBC.

Identifying relevant text spans is important for several downstream tasks in NLP, as it contributes to model explainability. While most span identification approaches rely on relatively smaller pre-trained language models like BERT, a few recent approaches have leveraged the latest generation of Large Language Models (LLMs) for the task. Current work has focused on explicit span identification like Named Entity Recognition (NER), while more subjective span identification with LLMs in tasks like Aspect-based Sentiment Analysis (ABSA) has been underexplored. In this paper, we fill this important gap by presenting an evaluation of the performance of various LLMs on text span identification in three popular tasks, namely sentiment analysis, offensive language identification, and claim verification. We explore several LLM strategies like instruction tuning, in-context learning, and chain of thought. Our results indicate underlying relationships within text aid LLMs in identifying precise text spans.

Ticket troubleshooting refers to the process of analyzing and resolving problems that are reported through a ticketing system. In large organizations offering a wide range of services, this task is highly complex due to the diversity of submitted tickets and the need for specialized domain knowledge. In particular, troubleshooting in telecommunications (telecom) is a very time-consuming task as it requires experts to interpret ticket content, consult documentation, and search historical records to identify appropriate resolutions. This human-intensive approach not only delays issue resolution but also hinders overall operational efficiency. To enhance the effectiveness and efficiency of ticket troubleshooting in telecom, we propose TeleDoCTR, a novel telecom-related, domain-specific, and contextual troubleshooting system tailored for end-to-end ticket resolution in telecom. TeleDoCTR integrates both domain-specific ranking and generative models to automate key steps of the troubleshooting workflow which are: routing tickets to the appropriate expert team responsible for resolving the ticket (classification task), retrieving contextually and semantically similar historical tickets (retrieval task), and generating a detailed fault analysis report outlining the issue, root cause, and potential solutions (generation task). We evaluate TeleDoCTR on a real-world dataset from a telecom infrastructure and demonstrate that it achieves superior performance over existing state-of-the-art methods, significantly enhancing the accuracy and efficiency of the troubleshooting process.

Language model (LM) probability is not a reliable quality estimator, as natural language is ambiguous. When multiple output options are valid, the model's probability distribution is spread across them, which can misleadingly indicate low output quality. This issue is caused by two reasons: (1) LMs' final output activation is softmax, which does not allow multiple correct options to receive high probabilities simultaneuously and (2) LMs' training data is single, one-hot encoded references, indicating that there is only one correct option at each output step. We propose training a module for Quality Estimation on top of pre-trained LMs to address these limitations. The module, called Sigmoid Head, is an extra unembedding head with sigmoid activation to tackle the first limitation. To tackle the second limitation, during the negative sampling process to train the Sigmoid Head, we use a heuristic to avoid selecting potentially alternative correct tokens. Our Sigmoid Head is computationally efficient during training and inference. The probability from Sigmoid Head is notably better quality signal compared to the original softmax head. As the Sigmoid Head does not rely on human-annotated quality data, it is more robust to out-of-domain settings compared to supervised QE.

Sequence modeling layers in modern language models typically face a trade-off between storage capacity and computational efficiency. While softmax attention offers unbounded storage at prohibitive quadratic cost, linear variants are more efficient but suffer from limited, fixed-size storage. We introduce Fast-weight Product Key Memory (FwPKM), a sparse fast-weight memory layer that resolves this tension. FwPKM updates sparsely activated parameters at both training and inference time using chunk-level gradient descent on a local memory-rewrite objective. This performs Test-Time Training (TTT)-style gradient updates on activated slots in a sparse memory, enabling rapid memorization and retrieval of many new key-value associations while keeping per-token compute low and fixed. Experiments show that FwPKM functions as an effective episodic memory that complements the semantic memory of standard modules, yielding significant perplexity reductions on long-context datasets. Notably, in Needle-in-a-Haystack evaluations, FwPKM generalizes to 128K-token contexts despite being trained on only 4K-token sequences.

Large Protein Language Models have shown strong potential for generative protein design, yet they frequently produce structural hallucinations, generating sequences with high linguistic likelihood that fold into thermodynamically unstable conformations. Existing alignment approaches such as Direct Preference Optimization are limited in this setting, as they model preferences as binary labels and ignore the continuous structure of the physical energy landscape. We propose Physio-DPO, a physics informed alignment framework that grounds protein language models in thermodynamic stability. Physio-DPO introduces a magnitude aware objective that scales optimization updates according to the energy gap between native structures and physics perturbed hard negatives. Experiments show that Physio-DPO consistently outperforms strong baselines including SFT, PPO, and standard DPO, reducing self consistency RMSD to 1.28 Å\ and increasing foldability to 92.8%. Qualitative analysis further demonstrates that Physio-DPO effectively mitigates structural hallucinations by recovering biophysical interactions such as hydrophobic core packing and hydrogen bond networks.

Large language models (LLMs) frequently produce contextual hallucinations, where generated content contradicts or ignores information explicitly stated in the prompt. Such errors are particularly problematic in deterministic automation workflows, where inputs are fixed and correctness is unambiguous. We introduce a simple and model-agnostic framework that provides explicit probabilistic guarantees for reducing hallucinations in this setting. We formalize the notion of a specific task, defined by a fixed input and a deterministic correctness criterion, and show that issuing the same prompt in independent context windows yields an exponential reduction in the probability that all model outputs are incorrect. To identify a correct answer among repeated runs, we incorporate an LLM-as-a-judge and prove that the probability that the judged pipeline fails decays at a rate determined by the judge's true- and false-positive probabilities. When the judge is imperfect, we strengthen it through majority vote over independent judge calls, obtaining ensemble-level error rates that decrease exponentially in the number of votes. This yields an explicit bound on the probability that the pipeline selects a hallucinated answer. Experiments on controlled extraction tasks with synthetic noisy judges match these predictions exactly: pipeline failure decreases exponentially with the number of repetitions, and hallucination-selection decreases exponentially with the number of judges in the ensemble. Together, these results provide a lightweight, modular, and theoretically grounded method for driving hallucination probabilities arbitrarily low in fixed-input LLM workflows-without modifying model weights, decoding strategies, or prompt engineering.

Traditional customer support systems, such as Interactive Voice Response (IVR), rely on rigid scripts and lack the flexibility required for handling complex, policy-driven tasks. While large language model (LLM) agents offer a promising alternative, evaluating their ability to act in accordance with business rules and real-world support workflows remains an open challenge. Existing benchmarks primarily focus on tool usage or task completion, overlooking an agent's capacity to adhere to multi-step policies, navigate task dependencies, and remain robust to unpredictable user or environment behavior. In this work, we introduce JourneyBench, a benchmark designed to assess policy-aware agents in customer support. JourneyBench leverages graph representations to generate diverse, realistic support scenarios and proposes the User Journey Coverage Score, a novel metric to measure policy adherence. We evaluate multiple state-of-the-art LLMs using two agent designs: a Static-Prompt Agent (SPA) and a Dynamic-Prompt Agent (DPA) that explicitly models policy control. Across 703 conversations in three domains, we show that DPA significantly boosts policy adherence, even allowing smaller models like GPT-4o-mini to outperform more capable ones like GPT-4o. Our findings demonstrate the importance of structured orchestration and establish JourneyBench as a critical resource to advance AI-driven customer support beyond IVR-era limitations.

Large language models (LLMs) are increasingly deployed in cost-sensitive and on-device scenarios, and safety guardrails have advanced mainly in English. However, real-world Chinese malicious queries typically conceal intent via homophones, pinyin, symbol-based splitting, and other Chinese-specific patterns. These Chinese-specific adversarial patterns create the safety evaluation gap that is not well captured by existing benchmarks focused on English. This gap is particularly concerning for lightweight models, which may be more vulnerable to such specific adversarial perturbations. To bridge this gap, we introduce the Chinese-Specific Safety Benchmark (CSSBench) that emphasizes these adversarial patterns and evaluates the safety of lightweight LLMs in Chinese. Our benchmark covers six domains that are common in real Chinese scenarios, including illegal activities and compliance, privacy leakage, health and medical misinformation, fraud and hate, adult content, and public and political safety, and organizes queries into multiple task types. We evaluate a set of popular lightweight LLMs and measure over-refusal behavior to assess safety-induced performance degradation. Our results show that the Chinese-specific adversarial pattern is a critical challenge for lightweight LLMs. This benchmark offers a comprehensive evaluation of LLM safety in Chinese, assisting robust deployments in practice.

Large language models (LLMs) have demonstrated significant advancements in reasoning and code generation, but efficiently creating new benchmarks to evaluate these capabilities remains a challenge. Traditional benchmark creation relies on manual human effort, which is expensive and time-consuming. Furthermore, existing benchmarks often contaminate LLM training data, necessitating novel and diverse benchmarks to accurately assess their genuine capabilities. This work introduces InfoSynth, a novel framework for automatically generating and evaluating reasoning benchmarks guided by information-theoretic principles. We propose metrics based on KL-divergence and entropy to quantify benchmark novelty and diversity without relying on costly model evaluations. Building on this framework, we develop an end-to-end pipeline that synthesizes robust Python coding problems from seed datasets using genetic algorithms and iterative code feedback. Our method generates accurate test cases and solutions to new problems 97% of the time, and the synthesized benchmarks consistently exhibit higher difficulty compared to prior works. Moreover, our algorithm provides a method for controlling the novelty/diversity and difficulty of generated problems. InfoSynth offers a scalable, self-verifying pipeline for constructing high-quality, challenging coding benchmarks for LLMs. Project Page: https://ishirgarg.github.io/infosynth_web/

Large-scale multilingual ASR (mASR) models such as Whisper achieve strong performance but incur high computational and latency costs, limiting their deployment on resource-constrained edge devices. In this study, we propose a lightweight and language-agnostic multilingual ASR system based on a CTC architecture with domain adaptation. Specifically, we introduce a Language-agnostic Hierarchical LoRA-MoE (HLoRA) framework integrated into an mHuBERT-CTC model, enabling end-to-end decoding via LID-posterior-driven LoRA routing. The hierarchical design consists of a multilingual shared LoRA for learning language-invariant acoustic representations and language-specific LoRA experts for modeling language-dependent characteristics. The proposed routing mechanism removes the need for prior language identity information or explicit language labels during inference, achieving true language-agnostic decoding. Experiments on MSR-86K and the MLC-SLM 2025 Challenge datasets demonstrate that HLoRA achieves comparable performance to two-stage inference approaches while reducing RTF by 11.7% and 8.2%, respectively, leading to improved decoding efficiency for low-resource mASR applications.

Language models are increasingly deployed under tight memory and compute constraints, where semantic routes (e.g., language, tone, or intent) may be assigned before decoding. However, these routes may not remain reflected in generated responses. We study this phenomenon as persistent semantic routing. We investigate this property using Embedding Consistency Regulation (ECR) as a minimal probe. ECR derives a short discrete prefix for each input using a frozen embedding model and applies the same prefix construction during both training and inference. In controlled prefix-assignment experiments, ECR achieves 99.93% $\pm$ 0.12% language routing accuracy, while random, shuffled, and dummy controls remain near chance (34–36%). This gap is not reliably captured by likelihood: some controls achieve competitive NLL despite losing routing consistency. We introduce Response Routing Recoverability (RRR), an encoder-independent diagnostic that tests whether assigned routes remain observable after generation. Routing recoverability persists across multilingual settings, unsupervised partitions, reduced precision, and transfer tasks. These findings establish routing persistence as a distinct and measurable property of language models, complementary to task performance and likelihood.

Multi-hop question answering (QA) requires systems to iteratively retrieve evidence and reason across multiple hops. While recent RAG and agentic methods report strong results, the underlying retrieval–reasoning process is often left implicit, making procedural choices hard to compare across model families. This survey takes the execution procedure as the unit of analysis and introduces a four-axis framework covering (A) overall execution plan, (B) index structure, (C) next-step control (strategies and triggers), and (D) stop/continue criteria. Using this schema, we map representative multi-hop QA systems and synthesize reported ablations and tendencies on standard benchmarks (e.g., HotpotQA, 2WikiMultiHopQA, MuSiQue), highlighting recurring trade-offs among effectiveness, efficiency, and evidence faithfulness. We conclude with open challenges for retrieval–reasoning agents, including structure-aware planning, transferable control policies, and robust stopping under distribution shift.

Do reasoning models have "Aha!" moments? Prior work suggests that models like DeepSeek-R1-Zero undergo sudden mid-trace realizations that lead to accurate outputs, implying an intrinsic capacity for self-correction. Yet, it remains unclear whether such intrinsic shifts in reasoning strategy actually improve performance. Here, we study mid-reasoning shifts and instrument training runs to detect them. Our analysis spans 1M+ reasoning traces, hundreds of training checkpoints, three reasoning domains, and multiple decoding temperatures and model architectures. We find that reasoning shifts are rare, do not become more frequent with training, and seldom improve accuracy, indicating that they do not correspond to prior perceptions of model insight. However, their effect varies with model uncertainty. Building on this finding, we show that artificially triggering extrinsic shifts under high entropy reliably improves accuracy. Our results show that mid-reasoning shifts are symptoms of unstable inference behavior rather than an intrinsic mechanism for self-correction.

Context engineering for large language model (LLM) agents requires distinguishing pragmatically useful information from misleading distractors. We introduce Entropic Context Shaping (ECS), an information-theoretic framework that measures context utility via the shift in the model's answer distribution toward the correct answer. Unlike lexical similarity methods that rely on word overlap, ECS captures pragmatic utility – whether a passage actually helps answer the question. We formalize utility as the signed change in answer probability and provide theoretical analysis showing that task-irrelevant updates yield near-zero distribution shift. We evaluate on multi-turn context selection tasks using LongMemEval (session-level) and LoCoMo (turn-level) benchmarks. On fine-grained turn selection, ECS with Llama-3.1-8B achieves F1=0.265, a 71.83% relative improvement over TF-IDF (F1=0.154), demonstrating that pragmatic utility outperforms lexical similarity when precise context selection matters. Code and data are available in the supplementary materials.

`SciHigh: Research Highlight Generation from Scientific Papers' focuses on the task of automatically generating concise, informative, and meaningful bullet-point highlights directly from scientific abstracts. The goal of this task is to evaluate how effectively computational models can generate highlights that capture the key contributions, findings, and novelty of a paper in a concise form. Highlights help readers grasp essential ideas quickly and are often easier to read and understand than longer paragraphs, especially on mobile devices. The track uses the MixSub dataset \cite{10172215}, which provides pairs of abstracts and corresponding author-written highlights. In this inaugural edition of the track, 12 teams participated, exploring various approaches, including pre-trained language models, to generate highlights from this scientific dataset. All submissions were evaluated using established metrics such as ROUGE, METEOR, and BERTScore to measure both alignment with author-written highlights and overall informativeness. Teams were ranked based on ROUGE-L scores. The findings suggest that automatically generated highlights can reduce reading effort, accelerate literature reviews, and enhance metadata for digital libraries and academic search platforms. SciHigh provides a dedicated benchmark for advancing methods aimed at concise and accurate highlight generation from scientific writing.

Question-answering (QA) models have advanced significantly in machine reading comprehension but often exhibit biases that hinder their performance, particularly with complex queries in adversarial conditions. This study evaluates the ELECTRA-small model on the Stanford Question Answering Dataset (SQuAD) v1.1 and adversarial datasets AddSent and AddOneSent. By identifying errors related to lexical bias, numerical reasoning, and entity recognition, we develop a multi-domain debiasing framework incorporating knowledge distillation, debiasing techniques, and domain expansion. Our results demonstrate up to 2.6 percentage point improvements in Exact Match (EM) and F1 scores across all test sets, with gains in adversarial contexts. These findings highlight the potential of targeted bias mitigation strategies to enhance the robustness and reliability of natural language understanding systems.

The Transformer architecture, a cornerstone of modern Large Language Models (LLMs), has achieved extraordinary success in sequence modeling, primarily due to its attention mechanism. However, despite its power, the standard attention mechanism is plagued by well-documented issues: representational collapse and attention sink. Although prior work has proposed approaches for these issues, they are often studied in isolation, obscuring their deeper connection. In this paper, we present a unified perspective, arguing that both can be traced to a common root – improper attention allocation. We identify two failure modes: 1) Attention Overload, where tokens receive comparable high weights, blurring semantic features that lead to representational collapse; 2) Attention Underload, where no token is semantically relevant, yet attention is still forced to distribute, resulting in spurious focus such as attention sink. Building on this insight, we introduce Lazy Attention, a novel mechanism designed for a more focused attention distribution. To mitigate overload, it employs positional discrimination across both heads and dimensions to sharpen token distinctions. To counteract underload, it incorporates Elastic-Softmax, a modified normalization function that relaxes the standard softmax constraint to suppress attention on irrelevant tokens. Experiments on the FineWeb-Edu corpus, evaluated across nine diverse benchmarks, demonstrate that Lazy Attention successfully mitigates attention sink and achieves competitive performance compared to both standard attention and modern architectures, while reaching up to 59.58% attention sparsity.

Search in e-Commerce is powered at the core by a structured representation of the inventory, often formulated as a category taxonomy. An important capability in e-Commerce with hierarchical taxonomies is to select a set of relevant leaf categories that are semantically aligned with a given user query. In this scope, we address a fundamental problem of search query categorization in real-world e-Commerce taxonomies. A correct categorization of a query not only provides a way to zoom into the correct inventory space, but opens the door to multiple intent understanding capabilities for a query. A practical and accurate solution to this problem has many applications in e-commerce, including constraining retrieved items and improving the relevance of the search results. For this task, we explore a novel Chain-of-Thought (CoT) paradigm that combines simple tree-search with LLM semantic scoring. Assessing its classification performance on human-judged query-category pairs, relevance tests, and LLM-based reference methods, we find that the CoT approach performs better than a benchmark that uses embedding-based query category predictions. We show how the CoT approach can detect problems within a hierarchical taxonomy. Finally, we also propose LLM-based approaches for query-categorization of the same spirit, but which scale better at the range of millions of queries.

Natural language often combines multiple ideas into complex sentences. Atomic sentence extraction, the task of decomposing complex sentences into simpler sentences that each express a single idea, improves performance in information retrieval, question answering, and automated reasoning systems. Previous work has formalized the "split-and-rephrase" task and established evaluation metrics, and machine learning approaches using large language models have improved extraction accuracy. However, these methods lack interpretability and provide limited insight into which linguistic structures cause extraction failures. Although some studies have explored dependency-based extraction of subject-verb-object triples and clauses, no principled analysis has examined which specific clause structures and dependencies lead to extraction difficulties. This study addresses this gap by analyzing how complex sentence structures, including relative clauses, adverbial clauses, coordination patterns, and passive constructions, affect the performance of rule-based atomic sentence extraction. Using the WikiSplit dataset, we implemented dependency-based extraction rules in spaCy, generated 100 gold=standard atomic sentence sets, and evaluated performance using ROUGE and BERTScore. The system achieved ROUGE-1 F1 = 0.6714, ROUGE-2 F1 = 0.478, ROUGE-L F1 = 0.650, and BERTScore F1 = 0.5898, indicating moderate-to-high lexical, structural, and semantic alignment. Challenging structures included relative clauses, appositions, coordinated predicates, adverbial clauses, and passive constructions. Overall, rule-based extraction is reasonably accurate but sensitive to syntactic complexity.