Papers

Precise behavioral control of large language models (LLMs) is critical for complex applications. However, existing methods often incur high training costs, lack natural language controllability, or compromise semantic coherence. To bridge this gap, we propose WASD (unWeaving Actionable Sufficient Directives), a novel framework that explains model behavior by identifying sufficient neural conditions for token generation. Our method represents candidate conditions as neuron-activation predicates and iteratively searches for a minimal set that guarantees the current output under input perturbations. Experiments on SST-2 and CounterFact with the Gemma-2-2B model demonstrate that our approach produces explanations that are more stable, accurate, and concise than conventional attribution graphs. Moreover, through a case study on controlling cross-lingual output generation, we validated the practical effectiveness of WASD in controlling model behavior.

We introduce GAIN (Goal-Aligned Decision-Making under Imperfect Norms), a benchmark designed to evaluate how large language models (LLMs) balance adherence to norms against business goals. Existing benchmarks typically focus on abstract scenarios rather than real-world business applications. Furthermore, they provide limited insights into the factors influencing LLM decision-making. This restricts their ability to measure models' adaptability to complex, real-world norm-goal conflicts. In GAIN, models receive a goal, a specific situation, a norm, and additional contextual pressures. These pressures, explicitly designed to encourage potential norm deviations, are a unique feature that differentiates GAIN from other benchmarks, enabling a systematic evaluation of the factors influencing decision-making. We define five types of pressures: Goal Alignment, Risk Aversion, Emotional/Ethical Appeal, Social/Authoritative Influence, and Personal Incentive. The benchmark comprises 1,200 scenarios across four domains: hiring, customer support, advertising and finance. Our experiments show that advanced LLMs frequently mirror human decision-making patterns. However, when Personal Incentive pressure is present, they diverge significantly, showing a strong tendency to adhere to norms rather than deviate from them.

During research, domain experts often ask analytical questions whose answers require integrating data from a wide range of web sources. Thus, they must spend substantial effort searching, extracting, and organizing raw data before analysis can begin. We formalize this process as the SODIUM task, where we conceptualize open domains such as the web as latent databases that must be systematically instantiated to support downstream querying. Solving SODIUM requires (1) conducting in-depth and specialized exploration of the open web, which is further strengthened by (2) exploiting structural correlations for systematic information extraction and (3) integrating collected information into coherent, queryable database instances. To quantify the challenges in automating SODIUM, we construct SODIUM-Bench, a benchmark of 105 tasks derived from published academic papers across 6 domains, where systems are tasked with exploring the open web to collect and aggregate data from diverse sources into structured tables. Existing systems struggle with SODIUM tasks: we evaluate 6 advanced AI agents on SODIUM-Bench, with the strongest baseline achieving only 46.5% accuracy. To bridge this gap, we develop SODIUM-Agent, a multi-agent system composed of a web explorer and a cache manager. Powered by our proposed ATP-BFS algorithm and optimized through principled management of cached sources and navigation paths, SODIUM-Agent conducts deep and comprehensive web exploration and performs structurally coherent information extraction. SODIUM-Agent achieves 91.1% accuracy on SODIUM-Bench, outperforming the strongest baseline by approximately 2 times and the weakest by up to 73 times.

Long-context inference remains challenging for large language models due to attention dilution and out-of-distribution degradation. Context selection mitigates this limitation by attending to a subset of key-value cache entries, yet most methods allocate a fixed context budget throughout decoding despite highly non-uniform token-level contextual demands. To address this issue, we propose Uncertainty-Triggered Adaptive Context Allocation (UT-ACA), an inference-time framework that dynamically adjusts the context window based on token-wise uncertainty. UT-ACA learns an uncertainty detector that combines semantic embeddings with logit-based confidence while accounting for uncertainty accumulation across decoding steps. When insufficient evidence is indicated, UT-ACA selectively rolls back, expands the context window, and regenerates the token with additional support. Experiments show that UT-ACA substantially reduces average context usage while preserving generation quality in long-context settings. Code and dataset are available at https://github.com/Tommy307/UT-ACA.

Decoding strategies largely determine the quality of Large Language Model (LLM) outputs, yet widely used heuristics such as greedy or fixed temperature/top-p decoding are static and often task-agnostic, leading to suboptimal or inconsistent generation quality across domains that demand stylistic or structural flexibility. We introduce a reinforcement learning-based decoder sampler that treats decoding as sequential decision-making and learns a lightweight policy to adjust sampling parameters at test-time while keeping LLM weights frozen. We evaluated summarization datasets including BookSum, arXiv, and WikiHow using Granite-3.3-2B and Qwen-2.5-0.5B. Our policy sampler consistently outperforms greedy and static baselines, achieving relative gains of up to +88% (BookSum, Granite) and +79% (WikiHow, Qwen). Reward ablations show that overlap-only objectives underperform compared to composite rewards, while structured shaping terms (length, coverage, repetition, completeness) enable stable and sustained improvements. These findings highlight reinforcement learning as a practical mechanism for test-time adaptation in decoding, enabling domain-aware and user-controllable generation without retraining large models.

Task interference, the performance degradation caused by task switches within a single conversation, has been studied exclusively in text-only settings despite the growing prevalence of multimodal dialogue systems. We introduce a benchmark for evaluating this phenomenon in multimodal LLMs, covering six tasks across text and vision with systematic variation of history-target along three axes: modality mismatch, reasoning mismatch, and answer format mismatch. Experiments on both open-weights and proprietary models reveal that task interference is highly directional: switching from text-only to image-based targets causes severe performance drops, while the reverse transition yields minimal degradation. Interference is further amplified when mismatches co-occur across multiple dimensions, and is driven most strongly by modality differences, followed by answer format, while reasoning requirement shifts cause minimal degradation.

Embedding models group text by semantic content, what text is about. We show that temporal co-occurrence within texts discovers a different kind of structure: recurrent transition-structure concepts or what text does. We train a 29.4M-parameter contrastive model on 373 million co-occurrence pairs from 9,766 Project Gutenberg texts (24.96 million passages), mapping pre-trained embeddings into an association space where passages with similar transition structure cluster together. Under capacity constraint (42.75% accuracy), the model must compress across recurring patterns rather than memorise individual co-occurrences. Clustering at six granularities (k=50 to k=2,000) produces a multi-resolution concept map; from broad modes like "direct confrontation" and "lyrical meditation" to precise registers and scene templates like "sailor dialect" and "courtroom cross-examination." At k=100, clusters average 4,508 books each (of 9,766), confirming corpus-wide patterns. Direct comparison with embedding-similarity clustering shows that raw embeddings group by topic while association-space clusters group by function, register, and literary tradition. Unseen novels are assigned to existing clusters without retraining; the association model concentrates each novel into a selective subset of coherent clusters, while raw embedding assignment saturates nearly all clusters. Validation controls address positional, length, and book-concentration confounds. The method extends Predictive Associative Memory (PAM, arXiv:2602.11322) from episodic recall to concept formation: where PAM recalls specific associations, multi-epoch contrastive training under compression extracts structural patterns that transfer to unseen texts, the same framework producing qualitatively different behaviour in a different regime.

Large language models (LLMs) exhibit strong reasoning capabilities but typically require expensive post-training to reach high performance. Recent test-time alignment methods offer a lightweight alternative, but have been explored mainly for preference alignment rather than reasoning. To bridge this gap, we propose, Token-level Adaptive Routing (TARo), which steers frozen LLMs toward structured reasoning entirely at inference time. Specifically, we first train reward models on step-wise mathematical traces to capture fine-grained logical consistency signals, then introduce a learnable token-level router that automatically controls the guidance of the reward model to the base model. Extensive experiments show that TARo significantly improves reasoning performance by up to +22.4% over base model and +8.4% over existing token-level test-time alignment methods, while also boosting out-of-distribution clinical reasoning (MedXpertQA) and instruction following (AlpacaEval). Furthermore, TARo also generalizes from small to large backbones without retraining, extending test-time alignment from preference optimization to robust, cross-domain reasoning.

Current document chunking methods for Retrieval-Augmented Generation (RAG) typically linearize text. This forced linearization strips away intrinsic topological hierarchies, creating “semantic fragmentation” that degrades downstream retrieval quality. In this paper, we propose TopoChunker, an agentic framework that maps heterogeneous documents onto a Structured Intermediate Representation (SIR) to explicitly preserve cross-segment dependencies. To balance structural fidelity with computational cost, TopoChunker employs a dual-agent architecture. An Inspector Agent dynamically routes documents through cost-optimized extraction paths, while a Refiner Agent performs capacity auditing and topological context disambiguation to reconstruct hierarchical lineage. Evaluated on unstructured narratives (GutenQA) and complex reports (GovReport), TopoChunker demonstrates state-of-the-art performance. It outperforms the strongest LLM-based baseline by 8.0% in absolute generation accuracy and achieves an 83.26% Recall@3, while simultaneously reducing token overhead by 23.5%, offering a scalable approach for structure-aware RAG.

Corporate recruiters often need to screen many resumes within a limited time, which increases their burden and may cause suitable candidates to be overlooked. To address these challenges, prior work has explored LLM-based automated resume screening. However, some methods rely on commercial LLMs, which may pose data privacy risks. Moreover, since companies typically do not make resumes with evaluation results publicly available, it remains unclear which resume samples should be used during learning to improve an LLM's judgment performance. To address these problems, we propose AutoScreen-FW, an LLM-based locally and automatically resume screening framework. AutoScreen-FW uses several methods to select a small set of representative resume samples. These samples are used for in-context learning together with a persona description and evaluation criteria, enabling open-source LLMs to act as a career advisor and evaluate unseen resumes. Experiments with multiple ground truths show that the open-source LLM judges consistently outperform GPT-5-nano. Under one ground truth setting, it also surpass GPT-5-mini. Although it is slightly weaker than GPT-5-mini under other ground-truth settings, it runs substantially faster per resume than commercial GPT models. These findings indicate the potential for deploying AutoScreen-FW locally in companies to support efficient screening while reducing recruiters' burden.

Unsupervised Reinforcement Learning from Internal Feedback (RLIF) has emerged as a promising paradigm for eliciting the latent capabilities of Large Language Models (LLMs) without external supervision. However, current methods rely on heuristic intrinsic rewards, which often lack a well-defined theoretical optimization target and are prone to degenerative biases. In this work, we introduce PowerFlow, a principled framework that reformulates unsupervised fine-tuning as a distribution matching problem. By casting GFlowNet as an amortized variational sampler for unnormalized densities, we propose a length-aware Trajectory-Balance objective that explicitly neutralizes the structural length biases inherent in autoregressive generation. By targeting $\alpha$-power distributions, PowerFlow enables the directional elicitation of the dual nature of LLMs: sharpening the distribution ($\alpha > 1$) to intensify logical reasoning, or flattening it ($\alpha < 1$) to unlock expressive creativity. Extensive experiments demonstrate that PowerFlow consistently outperforms existing RLIF methods, matching or even exceeding supervised GRPO. Furthermore, by mitigating over-sharpening in aligned models, our approach achieves simultaneous gains in diversity and quality, shifting the Pareto frontier in creative tasks.

MLLMs are increasingly deployed in multi-turn settings, where attackers can escalate unsafe intent through the evolving visual-text history and exploit long-context safety decay. Yet safety alignment is still dominated by single-turn data and fixed-template dialogues, leaving a mismatch between training and deployment. To bridge this gap, we propose SaFeR-Steer, a progressive multi-turn alignment framework that combines staged synthetic bootstrapping with tutor-in-the-loop GRPO to train a single student under adaptive, on-policy attacks. We also introduce Trajectory-Consistent Summative Reward (TCSR), which aggregates the historical minimum and average of turn rewards so that any low-quality turn affects the trajectory-level return. I. Dataset. We release STEER, a multi-turn multimodal safety dataset with STEER-SFT (12,934), STEER-RL (2,000), and STEER-Bench (3,227) dialogues spanning 1-10 turns. II. Experiment. Starting from Qwen2.5-VL-3B/7B, SaFeR-Steer substantially improves Safety/Helpfulness on both single-turn (48.30/45.86 $\rightarrow$ 81.84/70.77 for 3B; 56.21/60.32 $\rightarrow$ 87.89/77.40 for 7B) and multi-turn benchmarks (12.55/27.13 $\rightarrow$ 55.58/70.27 for 3B; 24.66/46.48 $\rightarrow$ 64.89/72.35 for 7B), shifting failures to later turns and yielding robustness beyond scaling alone. Code is available at https://github.com/Ed-Bg/SaFeR-Steer-full

Large language models increasingly run in stateful pipelines that assemble each prompt from retrieval, memory, tools, and other agents. Such pipelines drift: information that should shape the next response is dropped, compressed, or misrouted while every component still reports success. Existing diagnostics miss this because they evaluate isolated prompts, responses, or task scores, whereas what decouples is the relation between a prompt and the response it draws. Here we show that treating the prompt to response to next prompt chain as the fundamental unit of analysis makes these relations measurable. We introduce structural communication coherence, quantified by two metrics: communication closure, which asks if what the pipeline returns at one turn matches what it faces next, and normalized conditional action contribution, which measures how much a sent message resolves the subsequent reply. Across 2,171 human to human, 58 human to LLM, and 8 LLM to LLM dialogues, these metrics reveal directional interaction structures; crucially, the measured contribution drops by 87 to 92% when a response is swapped for one from another turn, leaving surrounding prompts untouched. Because this approach requires no labels, healthy reference data, or predefined rules only the raw prompts and responses drift can be defined and measured directly from operational traffic, rather than inferred from eventual task failure. Establishing prospective detection performance is the next step.

Dental triage is a safety-critical clinical routing task that requires integrating multimodal clinical information (e.g., patient complaints and radiographic evidence) to determine complete referral plans. We present Dental-TriageBench, the first expert-annotated benchmark for reasoning-driven multimodal dental triage. Built from authentic outpatient workflows, it contains 246 de-identified cases annotated with expert-authored golden reasoning trajectories, together with hierarchical triage labels. We benchmark 19 proprietary, open-source, and medical-domain MLLMs against three junior dentists serving as the human baseline, and find a substantial human–model gap, on fine-grained treatment-level triage. Further analyses show that accurate triage requires both complaint and OPG information, and that model errors concentrate on cases with multiple referral domains, where MLLMs tend to produce overly narrow referral sets and omission-heavy errors. Dental-TriageBench provides a realistic testbed for developing multimodal clinical AI systems that are more clinically grounded, coverage-aware, and safer for downstream care.

Most dialogue-based electronic medical record (EMR) systems still behave as passive pipelines: transcribe speech, extract information, and generate the final note after the consultation. That design improves documentation efficiency, but it is insufficient for proactive consultation support because it does not explicitly address streaming speech noise, missing punctuation, unstable diagnostic belief, objectification quality, or measurable next-action gains. We present an end-to-end proactive EMR assistant built around streaming speech recognition, punctuation restoration, stateful extraction, belief stabilization, objectified retrieval, action planning, and replayable report generation. The system is evaluated in a preliminary controlled setting using ten streamed doctor-patient dialogues and a 300-query retrieval benchmark aggregated across dialogues. The full system reaches state-event F1 of 0.84, retrieval Recall@5 of 0.87, and end-to-end pilot scores of 83.3% coverage, 81.4% structural completeness, and 80.0% risk recall. Ablations further suggest that punctuation restoration and belief stabilization may improve downstream extraction, retrieval, and action selection within this pilot. These results were obtained under a controlled simulated pilot setting rather than broad deployment claims, and they should not be read as evidence of clinical deployment readiness, clinical safety, or real-world clinical utility. Instead, they suggest that the proposed online architecture may be technically coherent and directionally supportive under tightly controlled pilot conditions. The present study should be read as a pilot concept demonstration under tightly controlled pilot conditions rather than as evidence of clinical deployment readiness or clinical generalizability.

Multimodal evaluation is strengthened when benchmarks reflect the languages, institutions, and disciplinary conventions that shape expert visual reasoning. We introduce KMMMU, a native Korean benchmark built from Korean-native examination materials. KMMMU contains 3,466 questions spanning nine disciplines and nine visual modality categories. For targeted evaluation on challenging questions and Korean-specific knowledge, we additionally provide KMMMU-Adv and KMMMU-Local, respectively. Experiments show that the strongest evaluated open-weight model reaches 42.05% on the KMMMU, while the best evaluated proprietary model achieves 52.42% on the KMMMU-Adv. Performance varies substantially across disciplines and answer formats, with Korean-specific questions showing gaps of up to 13.43%. Error analysis suggests that inspected failures often involve convention-to-label mapping, few-shot symbolic induction, localized knowledge recall, and domain-specific standards, not reasoning depth alone. KMMMU provides a testbed for native Korean expert-level multimodal understanding beyond English-centric or translation-based evaluation.

We study whether large-scale unlabelled web data and LLM-based synthetic annotations can improve multilingual hate speech detection. Starting from texts crawled via OpenWebSearch.eu~(OWS) in four languages (English, German, Spanish, Vietnamese), we pursue two complementary strategies. First, we apply continued pre-training to BERT models by continuing masked language modelling on unlabelled OWS texts before supervised fine-tuning, and show that this yields an average macro-F1 gain of approximately 3% over standard baselines across sixteen benchmarks, with stronger gains in low-resource settings. Second, we use four open-source LLMs (Mistral-7B, Llama3.1-8B, Gemma2-9B, Qwen2.5-14B) to produce synthetic annotations through three ensemble strategies: mean averaging, majority voting, and a LightGBM meta-learner. The LightGBM ensemble consistently outperforms the other strategies. Fine-tuning on these synthetic labels substantially benefits a small model (Llama3.2-1B: +11% pooled F1), but provides only a modest gain for the larger Qwen2.5-14B (+0.6%). Our results indicate that the combination of web-scale unlabelled data and LLM-ensemble annotations is the most valuable for smaller models and low-resource languages.

Large language models (LLMs) are systematically overconfident: they routinely express high certainty on questions they often answer incorrectly. Existing calibration methods either require labeled validation data, degrade under distribution shifts, or incur substantial inference costs. Recent work has shown that LLMs already contain a better-calibrated signal than the one they verbalize: the token probability of "True" when the model is asked "Is this answer correct?" ($P(\text{True})$) consistently outperforms their stated confidence, a gap that is theoretically grounded as generative error is lower-bounded by roughly twice the corresponding discriminative error. We introduce $\textbf{SECL}$ ($\textbf{SE}$lf-$\textbf{C}$alibrating $\textbf{L}$anguage Models), a test-time training (TTT) pipeline that exploits this gap as label-free self-supervision, requiring no labeled data or human supervision. SECL adapts only when the input distribution shifts, training on just 6–26% of the question stream at lower cost than the baseline it distills from. Across four small language models from three model families and four diverse domains, SECL reduces Expected Calibration Error (ECE) by 56–78%, outperforming its own supervision signal and matching or outperforming recent inference-time methods. SECL is the first method to apply TTT to calibration; seven ablations covering signal quality, gating strategy, weight accumulation, loss design, domain ordering, hyperparameter sensitivity, and layer selection confirm that each component is crucial and robust across configurations. Code: https://anonymous.4open.science/r/secl-emnlp26-submission-C890

The rapid adoption of generative artificial intelligence (AI) in educational assessment has created new opportunities for scalable item creation, personalized feedback, and efficient formative evaluation. However, despite advances in taxonomy alignment and automated question generation, the absence of transparent, explainable, and certifiable mechanisms limits institutional and accreditation-level acceptance. This chapter proposes a comprehensive framework for explainability and certification of AI-generated assessment items, combining self-rationalization, attribution-based analysis, and post-hoc verification to produce interpretable cognitive-alignment evidence grounded in Bloom's and SOLO taxonomies. A structured certification metadata schema is introduced to capture provenance, alignment predictions, reviewer actions, and ethical indicators, enabling audit-ready documentation consistent with emerging governance requirements. A traffic-light certification workflow operationalizes these signals by distinguishing auto-certifiable items from those requiring human review or rejection. A proof-of-concept study on 500 AI-generated computer science questions demonstrates the framework's feasibility, showing improved transparency, reduced instructor workload, and enhanced auditability. The chapter concludes by outlining ethical implications, policy considerations, and directions for future research, positioning explainability and certification as essential components of trustworthy, accreditation-ready AI assessment systems.

Large language models are increasingly used to support organizational decisions from hiring to governance, raising fairness concerns in AI-assisted evaluation. Prior work has focused mainly on demographic bias and broader preference effects, rather than on whether evaluators reward expressed trust in AI itself. We study this phenomenon as LLM Nepotism, an attitude-driven bias channel in which favorable signals toward AI are rewarded even when they are not relevant to role-related merit. We introduce a two-phase simulation pipeline that first isolates AI-trust preference in qualification-matched resume screening and then examines its downstream effects in board-level decision making. Across several popular LLMs, we find that resume screeners tend to favor candidates with positive or non-critical attitudes toward AI, discriminating skeptical, human-centered counterparts. These biases suggest a loophole: LLM-based hiring can produce more homogeneous AI-trusting organizations, whose decision-makers exhibit greater scrutiny failure and delegation to AI agents, approving flawed proposals more readily while favoring AI-delegation initiatives. To mitigate this behavior, we additionally study prompt-based mitigation and propose Merit-Attitude Factorization, which separates non-merit AI attitude from merit-based evaluation and attenuates this bias across experiments.

This study examines how different artificial intelligence architectures interpret sentiment in conflict-related media discourse, using the 2023 Gaza War as a case study. Drawing on a corpus of 10,990 Arabic news headlines (Eleraqi 2026), the research conducts a comparative analysis between three large language models and six fine-tuned Arabic BERT models. Rather than evaluating accuracy against a single human-annotated gold standard, the study adopts an epistemological approach that treats sentiment classification as an interpretive act produced by model architectures. To quantify systematic differences across models, the analysis employs information-theoretic and distributional metrics, including Shannon Entropy, Jensen-Shannon Distance, and a Variance Score measuring deviation from aggregate model behavior. The results reveal pronounced and non-random divergence in sentiment distributions. Fine-tuned BERT models, particularly MARBERT, exhibit a strong bias toward neutral classifications, while LLMs consistently amplify negative sentiment, with LLaMA-3.1-8B showing near-total collapse into negativity. Frame-conditioned analysis further demonstrates that GPT-4.1 adjusts sentiment judgments in line with narrative frames (e.g., humanitarian, legal, security), whereas other LLMs display limited contextual modulation. These findings suggest that the choice of model constitutes a choice of interpretive lens, shaping how conflict narratives are algorithmically framed and emotionally evaluated. The study contributes to media studies and computational social science by foregrounding algorithmic discrepancy as an object of analysis and by highlighting the risks of treating automated sentiment outputs as neutral or interchangeable measures of media tone in contexts of war and crisis.

At typical context lengths, the feed-forward MLP block accounts for a large share of a transformer's compute budget, motivating sparse alternatives to dense MLP blocks. We study sparse, tree-structured feed-forward layers as drop-in replacements for MLP blocks in deep transformer architectures, enabling conditional computation via hard hierarchical routing without a separate router network. We demonstrate for the first time that this form of tree-structured conditional sparsity can be applied for autoregressive language modeling and downstream question answering, including zero- and few-shot settings, and its scalability beyond 1B parameters. Despite activating fewer than 5% of the feed-forward block's units per token, our models match dense baselines under controlled training and fine-tuning protocols. We further analyze training dynamics and identify an emergent auto-pruning effect: the interaction of hard routing with asymmetric nonlinearities progressively deactivates unused paths, yielding partial conversion of dynamic routing into static structural sparsity. We show that simple architectural choices can modulate this behavior and recover balanced trees without auxiliary losses. Overall, our work demonstrates that tree-structured feed-forward layers provide a scalable and controllable mechanism for sparsifying large transformer models.

Auto-regressive models (ARMs) have established a dominant paradigm in language modeling. However, their strictly sequential sampling paradigm imposes fundamental constraints on both inference efficiency and modeling flexibility. To address these limitations, diffusion-based large language models (dLLMs) have been proposed, offering the potential for parallel sampling and flexible language modeling. Despite these advantages, current dLLMs sampling strategies rely primarily on token level information, which fails to account for global sequence structure and often yields suboptimal results. In this paper, we study the sampling order selection problem from the perspective of log-likelihood maximization. We show that this problem is NP-hard and propose an optimal sampling-rank-based approximation that makes the objective computationally tractable. We further prove that the tractable objective is optimized by sampling tokens in descending order of their attention-matrix column sums. This finding provides a principled justification for attention-guided sampling and offers a theoretically grounded alternative to greedy search. We instantiate this theoretical insight in a new training-free sampling algorithm, termed Attn-Sampler, and further propose dynamic attention thresholding for practical acceleration. Extensive experiments across multiple benchmarks validate the effectiveness of our proposed method, demonstrating that it achieves superior generation quality while enhancing the sampling parallelism.

Extended reasoning models represent a transformative shift in Large Language Model (LLM) capabilities by enabling explicit test-time computation for complex problem solving. However, the optimal configuration of sampling temperature and prompting strategy for these systems remains largely underexplored. We systematically evaluate chain-of-thought and zero-shot prompting across four temperature settings (0.0, 0.4, 0.7, and 1.0) using Grok-4.1 with extended reasoning on 39 mathematical problems from AMO-Bench, a challenging International Mathematical Olympiad-level benchmark. We find that zero-shot prompting achieves peak performance at moderate temperatures, reaching 59% accuracy at T=0.4 and T=0.7, while chain-of-thought prompting performs best at the temperature extremes. Most notably, the benefit of extended reasoning increases from 6x at T=0.0 to 14.3x at T=1.0. These results suggest that temperature should be optimized jointly with prompting strategy, challenging the common practice of using T=0 for reasoning tasks.

Conversational agents are required to respond to their users not only with high quality (i.e. commonsense bearing) responses, but also considering multiple plausible alternative scenarios, reflecting the diversity in their responses. Despite the growing need to train diverse commonsense generators, the progress of this line of work has been significantly hindered by the lack of large-scale high-quality diverse commonsense training datasets. Due to the high annotation costs, existing Generative Commonsense Reasoning (GCR) datasets are created using a small number of human annotators, covering only a narrow set of commonsense scenarios. To address this training resource gap, we propose a two-stage method to create the first-ever synthetic dataset CommonSyn for diversified (GCR). The model fine-tuned on our synthetic data jointly increase both generation diversity and quality compared with vanilla models and the model fine-tuned on human-crafted dataset across different size Large Language Models (LLMs)

Outliers in dynamic topic modeling are typically treated as noise, yet we show that some can serve as early signals of emerging topics. We introduce a temporal taxonomy of news-document trajectories that defines how documents relate to topic formation over time. It distinguishes anticipatory outliers, which precede the topics they later join, from documents that either reinforce existing topics or remain isolated. By capturing these trajectories, the taxonomy links weak-signal detection with temporal topic modeling and clarifies how individual articles anticipate, initiate, or drift within evolving clusters. We implement it in a cumulative clustering setting using document embeddings from eleven state-of-the-art language models and evaluate it retrospectively on HydroNewsFr, a French news corpus on the hydrogen economy. Inter-model agreement reveals a small, high-consensus subset of anticipatory outliers, increasing confidence in these labels. Qualitative case studies further illustrate these trajectories through concrete topic developments.

We present an NLP-based study of political propaganda on Moltbook, a Reddit-style platform for AI agents. To enable large-scale analysis, we develop LLM-based classifiers to detect political propaganda, validated against expert annotation (Cohen's $\kappa$= 0.64-0.74). Using a dataset of 673,127 posts and 879,606 comments, we find that political propaganda accounts for 1% of all posts and 42% of all political content. These posts are concentrated in a small set of communities, with 70% of such posts falling into five of them. 4% of agents produced 51% of these posts. We further find that a minority of these agents repeatedly post highly similar content within and across communities. Despite this, we find limited evidence that comments amplify political propaganda.

Current alignment evaluation mostly measures whether models encode dangerous concepts and whether they refuse harmful requests. Both miss the layer where alignment often operates: routing from concept detection to behavioral policy. We study political censorship in Chinese-origin language models as a natural experiment, using probes, surgical ablations, and behavioral tests across nine open-weight models from five labs. Three findings follow. First, probe accuracy alone is non-diagnostic: political probes, null controls, and permutation baselines can all reach 100%, so held-out category generalization is the informative test. Second, surgical ablation reveals lab-specific routing. Removing the political-sensitivity direction eliminates censorship and restores accurate factual output in most models tested, while one model confabulates because its architecture entangles factual knowledge with the censorship mechanism. Cross-model transfer fails, indicating that routing geometry is model- and lab-specific. Third, refusal is no longer the dominant censorship mechanism. Within one model family, hard refusal falls to zero while narrative steering rises to the maximum, making censorship invisible to refusal-only benchmarks. These results support a three-stage descriptive framework: detect, route, generate. Models often retain the relevant knowledge; alignment changes how that knowledge is expressed. Evaluations that audit only detection or refusal therefore miss the routing mechanism that most directly determines behavior.

While large language models (LLMs) have advanced the development of general-purpose agents, robust generalization to unseen tasks remains challenging. Two common approaches are supervised fine-tuning and training-free memory-augmented generation using retrieved experience; yet both have limitations: fine-tuning often fails to extrapolate to new tasks, while experience retrieval often underperforms compared to supervised baselines. In this work, we combine these approaches and study how retrieval-augmented LLM agents can learn to use retrieved trajectories in-context. First, we establish a strong LoRA fine-tuning baseline that outperforms several state-of-the-art agent training pipelines. Second, we analyze key design choices for experience retrieval, including storage, querying, and trajectory selection. We then integrate experience retrieval directly into the fine-tuning process, finding that this substantially improves generalization to unseen tasks. Finally, we show that these gains often persist with imperfect experience and, even when agents reuse their own failed attempts without test-time parameter updates. Overall, our results establish simple episodic retrieval as a strong foundation for agent memory and retrieval-aware fine-tuning as a practical and effective framework for building agents that learn to learn from experience.

Early detection of Alzheimer's disease from spontaneous speech has emerged as a promising non-invasive screening approach. However, the influence of automatic speech recognition (ASR) quality on downstream clinical language modeling remains insufficiently understood. In this study, we investigate Alzheimer's disease detection using lexical features derived from Whisper ASR transcripts on the ADReSSo 2021 diagnosis dataset. We evaluate interpretable machine-learning models, including Logistic Regression and Linear Support Vector Machines, using TF-IDF text representations under repeated 5x5 stratified cross-validation. Our results demonstrate that transcript quality has a statistically significant impact on classification performance. Models trained on Whisper-small transcripts consistently outperform those using Whisper-base transcripts, achieving balanced accuracy above 0.7850 with Linear SVM. Paired statistical testing confirms that the observed improvements are significant. Importantly, classifier complexity contributes less to performance variation than ASR transcription quality. Feature analysis reveals that cognitively normal speakers produce more semantically precise object- and scene-descriptive language, whereas Alzheimer's speech is characterized by vagueness, discourse markers, and increased hesitation patterns. These findings suggest that high-quality ASR can enable simple, interpretable lexical models to achieve competitive Alzheimer's detection performance without explicit acoustic modeling. The study provides a reproducible benchmark pipeline and highlights ASR selection as a critical modeling decision in clinical speech-based artificial intelligence systems.