Papers

Automatic Speech Recognition (ASR) systems are widely used in everyday communication, education, healthcare, and industry, yet their performance remains uneven across speakers, particularly when dialectal variation diverges from the mainstream accents represented in training data. This study investigates ASR bias through a sociolinguistic analysis of Newcastle English, a regional variety of North-East England that has been shown to challenge current speech recognition technologies. Using spontaneous speech from the Diachronic Electronic Corpus of Tyneside English (DECTE), we evaluate the output of a state-of-the-art commercial ASR system and conduct a fine-grained analysis of more than 3,000 transcription errors. Errors are classified by linguistic domain and examined in relation to social variables including gender, age, and socioeconomic status. In addition, an acoustic case study of selected vowel features demonstrates how gradient phonetic variation contributes directly to misrecognition. The results show that phonological variation accounts for the majority of errors, with recurrent failures linked to dialect-specific features like vowel quality and glottalisation, as well as local vocabulary and non-standard grammatical forms. Error rates also vary across social groups, with higher error frequencies observed for men and for speakers at the extremes of the age spectrum. These findings indicate that ASR errors are not random but socially patterned and can be explained from a sociolinguistic perspective. Thus, the study demonstrates the importance of incorporating sociolinguistic expertise into the evaluation and development of speech technologies and argues that more equitable ASR systems require explicit attention to dialectal variation and community-based speech data.

Activation steering has emerged as a powerful tool to shape LLM behavior without the need for weight updates. While its inherent brittleness and unreliability are well-documented, its safety implications remain underexplored. In this work, we present a systematic safety audit of steering vectors obtained with Contrastive Activation Addition (CAA), a widely used steering approach, under a unified evaluation protocol. Using JailbreakBench as benchmark, we show that steering vectors consistently influence the success rate of jailbreak attacks, with stronger amplification under simple template-based attacks. Across LLM families and sizes, steering the model in specific directions can drastically increase (up to 57%) or decrease (up to 50%) its attack success rate (ASR), depending on the targeted behavior. We attribute this phenomenon to the overlap between the steering vectors and the latent directions of refusal behavior. Thus, we offer a traceable explanation for this discovery. Together, our findings reveal the previously unobserved origin of this safety gap in LLMs, highlighting a trade-off between controllability and safety.

Reading comprehension presents a significant challenge for children with Special Educational Needs and Disabilities (SEND), often requiring intensive one-on-one reading support. To assist therapists in scaling this support, we developed a multilingual, AI-powered interface that automatically enhances text with visual scaffolding. This system dynamically identifies key concepts and maps them to contextually relevant pictograms, supporting learners across languages. We evaluated the system across five typologically diverse languages (English, French, Italian, Spanish, and Arabic), through multilingual coverage analysis, expert clinical review by speech therapists and special education professionals, and latency assessment. Evaluation results indicate high pictogram coverage and visual scaffolding density across the five languages. Expert audits suggested that automatically selected pictograms were semantically appropriate, with combined correct and acceptable ratings exceeding 95% for the four European languages and approximately 90% for Arabic despite reduced pictogram repository coverage. System latency remained within interactive thresholds suitable for real-time educational use. These findings support the technical viability, semantic safety, and acceptability of automated multimodal scaffolding to improve accessibility for neurodiverse learners.

Adaptive scaffolding enhances learning, yet the field lacks robust methods for measuring it within authentic tutoring dialogue. This gap has become more pressing with the rise of remote human tutoring and large language model-based systems. We introduce an embedding-based approach that analyzes scaffolding dynamics by aligning the semantics of dialogue turns, problem statements, and correct solutions. Specifically, we operationalize alignment by computing cosine similarity between tutor and student contributions and task-relevant content. We apply this framework to 1,576 real-world mathematics tutoring dialogues from the Eedi Question Anchored Tutoring Dialogues dataset. The analysis reveals systematic differences in task alignment and distinct temporal patterns in how participants ground their contributions in problem and solution content. Further, mixed-effects models show that role-specific semantic alignment predicts tutorial progression beyond baseline features such as message order and length. Tutor contributions exhibited stronger grounding in problem content early in interactions. In contrast, student solution alignment was modestly positively associated with progression. These findings support scaffolding as a continuous, role-sensitive process grounded in task semantics. By capturing role-specific alignment over time, this approach provides a principled method for analyzing instructional dialogue and evaluating conversational tutoring systems.

Miscalibrated confidence scores are a practical obstacle to deploying AI in clinical settings. A model that is always overconfident offers no useful signal for deferral. We present a multi-agent framework that combines domain-specific specialist agents with Two-Phase Verification (Wu et al., 2024) and S-Score Weighted Fusion to improve both calibration and discrimination in medical multiple-choice question answering. Four specialist agents (respiratory, cardiology, neurology, gastroenterology) generate independent diagnoses using Qwen2.5-7B-Instruct. Each diagnosis undergoes a two-phase self-verification process that measures internal consistency and produces a Specialist Confidence Score (S-score). The S-scores drive a weighted fusion strategy that selects the final answer and calibrates the reported confidence. We evaluate on high-disagreement subsets of MedQA-USMLE and MedMCQA (100 and 250 questions). All results are specific to this filtered regime. On MedQA-250, the full system achieves ECE = 0.091 (74.4% reduction over the single-specialist baseline) and AUROC = 0.630 (+0.056) at 59.2% accuracy. Calibration gains of 49-74% hold across all four settings. Ablation analysis reveals that Two-Phase Verification drives ECE reduction while multi-agent reasoning drives AUROC improvement, suggesting that consistency checking and ensemble aggregation address different failure modes of LLM uncertainty. Whether the resulting confidence signal is sufficient to support clinical deferral decisions in practice remains a direction for future investigation.

Self-distillation has emerged as an effective post-training paradigm for LLMs, often improving performance while shortening reasoning traces. However, in mathematical reasoning, we find that it can reduce response length while degrading performance. We trace this degradation to the suppression of epistemic verbalization - the model's expression of uncertainty during reasoning. Through controlled experiments varying conditioning context richness and task coverage, we show that conditioning the teacher on rich information suppresses uncertainty expression, enabling rapid in-domain optimization with limited task coverage but harming OOD performance, where unseen problems benefit from expressing uncertainty and adjusting accordingly. Across Qwen3-1.7B/8B, DeepSeek-Distill-Qwen-7B, and Olmo3-7B-Instruct, we observe performance drops of up to 40%. Our findings highlight that exposing appropriate levels of uncertainty is crucial for robust reasoning and underscore the importance of optimizing reasoning behavior beyond merely reinforcing correct answer traces.

Every year, 10 million pets enter shelters, separated from their families. Despite desperate searches by both guardians and lost animals, 70% never reunite, not because matches do not exist, but because current systems look only at appearance, while animals recognize each other through sound. We ask, why does computer vision treat vocalizing species as silent visual objects? Drawing on five decades of cognitive science showing that animals perceive quantity approximately and communicate identity acoustically, we present the first multimodal reunification system integrating visual and acoustic biometrics. Our species-adaptive architecture processes vocalizations from 10Hz elephant rumbles to 4kHz puppy whines, paired with probabilistic visual matching that tolerates stress-induced appearance changes. This work demonstrates that AI grounded in biological communication principles can serve vulnerable populations that lack human language.

Recent advances in large language models (LLMs) and LLM-based agents have substantially improved the capabilities of automated theorem proving. However, for problems that require complex mathematical reasoning, current systems seldom succeed in their initial attempt, necessitating iterative adjustments to their proof strategies. Existing approaches for handling failed attempts typically either iteratively fix errors within the proof or discard the entire proof and regenerate it from scratch. The former leads to progressively longer contexts, which degrade the model's ability to attend to the remaining unresolved subproblems, while the latter is inefficient, as it may abandon mostly correct reasoning due to localized errors. To address this dilemma, we present MechMath, an agent system centered on a Sorrifier-driven formal decomposition paradigm. By leveraging the sorry placeholder in Lean to precisely isolate unresolved subgoals while preserving the surrounding verified proof structure, MechMath extracts each failed subproblem into a clean, self-contained context and resolves it independently. This avoids both the waste of full regeneration and the excessive context length induced by repeated repairs. Experimental results on challenging mathematical competition benchmarks, including IMO 2025, Putnam 2025, miniF2F, and a subset of ProverBench, demonstrate that our agent achieves significant advantages in proving efficiency.

Speaker verification at large scale remains an open challenge as fixed-margin losses treat all samples equally regardless of quality. We hypothesize that mislabeled or degraded samples introduce noisy gradients that disrupt compact speaker manifolds. We propose Curry (CURriculum Ranking), an adaptive loss that estimates sample difficulty online via Sub-center ArcFace: confidence scores from dominant sub-center cosine similarity rank samples into easy, medium, and hard tiers using running batch statistics, without auxiliary annotations. Learnable weights guide the model from stable identity foundations through manifold refinement to boundary sharpening. To our knowledge, this is the largest-scale speaker verification system trained to date. Evaluated on VoxCeleb1-O, and SITW, Curry reduces EER by 86.8% and 60.0% over the Sub-center ArcFace baseline, establishing a new paradigm for robust speaker verification on imperfect large-scale data.

Generative Retrieval (GR) has emerged as a promising paradigm for modern search systems. Compared to multi-stage cascaded architecture, it offers advantages such as end-to-end joint optimization and high computational efficiency. OneSearch, as a representative industrial-scale deployed generative search framework, has brought significant commercial and operational benefits. However, its inadequate understanding of complex queries, inefficient exploitation of latent user intents, and overfitting to narrow historical preferences have limited its further performance improvement. To address these challenges, we propose OneSearch-V2, a latent reasoning enhanced self-distillation generative search framework. It contains three key innovations: (1) a thought-augmented complex query understanding module, which enables deep query understanding and overcomes the shallow semantic matching limitations of direct inference; (2) a reasoning-internalized self-distillation training pipeline, which uncovers users' potential yet precise e-commerce intentions beyond log-fitting through implicit in-context learning; (3) a behavior preference alignment optimization system, which mitigates reward hacking arising from the single conversion metric, and addresses personal preference via direct user feedback. Extensive offline evaluations demonstrate OneSearch-V2's strong query recognition and user profiling capabilities. Online A/B tests further validate its business effectiveness, yielding +3.98% item CTR, +2.07% buyer volume, and +2.11% order volume. Manual evaluation further confirms gains in search experience quality, with +1.37% in page good rate and +1.65% in query-item relevance. More importantly, OneSearch-V2 effectively mitigates common search system issues such as information bubbles and long-tail sparsity, without incurring additional inference costs or serving latency.

Poetry generation in Sanskrit typically requires the verse to be semantically coherent and adhere to strict prosodic rules. In Sanskrit prosody, every line of a verse is typically a fixed length sequence of syllables adhering to prescribed binary patterns of syllable weights. We observe that instead of treating a verse as a monolithic sequence, segmenting them as grouped-lines leads to significant improvement in semantic coherence by 10% with comparable metrical adherence. Specifically, Pingala, our proposed decoding approach is designed to encourage every line to have well-formed words and our token selection biases the model towards it by preferring longer tokens. Writing in Sanskrit follows phonemic orthography, hence using a phonetically aware transliteration scheme, SLP1, increased the metrical alignment by 46% with comparable semantic similarity, for a instruction fine-tuned large language models like Phi-4. We also introduce a new approach for reference-free evaluation using cross-encoders which achieved better alignment with true poetry instances.

High-quality teacher-child interaction (TCI) is fundamental to early childhood development, yet traditional expert-based assessment faces a critical scalability challenge. In large systems like China's-serving 36 million children across 250,000+ kindergartens-the cost and time requirements of manual observation make continuous quality monitoring infeasible, relegating assessment to infrequent episodic audits that limit timely intervention and improvement tracking. In this paper, we investigate whether AI can serve as a scalable assessment teammate by extracting structured quality indicators and validating their alignment with human expert judgments. Our contributions include: (1) TEPE-TCI-370h (Tracing Effective Preschool Education), the first large-scale dataset of naturalistic teacher-child interactions in Chinese preschools (370 hours, 105 classrooms) with standardized ECQRS-EC and SSTEW annotations; (2) We develop Interaction2Eval, a specialized LLM-based framework addressing domain-specific challenges-child speech recognition, Mandarin homophone disambiguation, and rubric-based reasoning-achieving up to 88% agreement; (3) Deployment validation across 43 classrooms demonstrating an 18x efficiency gain in the assessment workflow, highlighting its potential for shifting from annual expert audits to monthly AI-assisted monitoring with targeted human oversight. This work not only demonstrates the technical feasibility of scalable, AI-augmented quality assessment but also lays the foundation for a new paradigm in early childhood education-one where continuous, inclusive, AI-assisted evaluation becomes the engine of systemic improvement and equitable growth.

Evaluating the pedagogical quality of AI tutors remains challenging: standard NLG metrics do not determine whether responses identify mistakes, scaffold reasoning, or avoid revealing the answers. For the task of mistake remediation, we derive a hierarchy of pedagogical aspects from human pairwise preferences on MRBench, and synthesize minimally contrastive response pairs that differ along key aspects (e.g., mistake identification and location, targetedness, scaffolding, actionability, clarity, and coherence). We develop and release Bradley-Terry preference models trained on weighted-sum rankings that we automatically create from MRBench, synthetic pairs, and data combinations. Using only synthetic data, our best model reaches 0.69 pairwise accuracy on a human preference test, and combining weighted-sum data with targeted synthetic groups improves accuracy to 0.74, outperforming larger general-purpose reward models while using only a 0.5B-parameter backbone.

Autoformalization - automatically translating natural language mathematical texts into formal proof language such as Lean4 - can help accelerate AI-assisted mathematical research, be it via proof verification or proof search. I fine-tune Qwen3.5-2B with LoRA for natural language to Lean4 formalization on FineLeanCorpus and consider three training regimes: supervised fine-tuning (SFT) with curriculum learning (difficulty 1 to 10), SFT without curriculum ordering, and reinforcement learning using group relative policy optimization (GRPO) with a cycle consistency reward. Cycle consistency measures how well the meaning of a statement is preserved through a NL to Lean4 to NL' loop, computed as cosine similarity of off-the-shelf sentence embeddings. On an unseen subset of FineLeanCorpus (FLC) and on PutnamBench, RL substantially outperforms both SFT variants (mean cycle consistency 0.669 vs. 0.513 on FLC; 0.561 vs. 0.422 on PutnamBench), while increasing cross-entropy loss by only 0.011 nats, with minimal impact on formalization quality. Curriculum ordering provides no measurable benefit over shuffled training.

Multimodal LLMs are increasingly deployed as perceptual backbones for autonomous agents in 3D environments, from robotics to virtual worlds. These applications require agents to perceive rapid state changes, attribute actions to the correct entities, and reason about concurrent multi-agent behaviors from a first-person perspective, capabilities that existing benchmarks do not adequately evaluate. We introduce GameplayQA, a framework for evaluating agentic-centric perception and reasoning through video understanding. Specifically, we densely annotate multiplayer 3D gameplay videos at 1.22 labels/second, with time-synced, concurrent captions of states, actions, and events structured around a triadic system of Self, Other Agents, and the World, a natural decomposition for multi-agent environments. From these annotations, we refined 2.4K diagnostic QA pairs organized into three levels of cognitive complexity, accompanied by a structured distractor taxonomy that enables fine-grained analysis of where models hallucinate. Evaluation of frontier MLLMs reveals a substantial gap from human performance, with common failures in temporal and cross-video grounding, agent-role attribution, and handling the decision density of the game. We hope GameplayQA stimulates future research at the intersection of embodied AI, agentic perception, and world modeling.

We release Samasāmayik, a novel, meticulously curated, large-scale Hindi-Sanskrit corpus, comprising 92,196 parallel sentences. Unlike most data available in Sanskrit, which focuses on classical era text and poetry, this corpus aggregates data from diverse sources covering contemporary materials, including spoken tutorials, children's magazines, radio conversations, and instruction materials. We benchmark this new dataset by fine-tuning three complementary models - ByT5, NLLB and IndicTrans-v2, to demonstrate its utility. Our experiments demonstrate that models trained on the Samasamayik corpus achieve significant performance gains on in-domain test data, while achieving comparable performance on other widely used test sets, establishing a strong new performance baseline for contemporary Hindi-Sanskrit translation. Furthermore, a comparative analysis against existing corpora reveals minimal semantic and lexical overlap, confirming the novelty and non-redundancy of our dataset as a robust new resource for low-resource Indic language MT.

The ground state search problem is central to quantum computing, with applications spanning quantum chemistry, condensed matter physics, and optimization. The Variational Quantum Eigensolver (VQE) has shown promise for small systems but faces significant limitations. These include barren plateaus, restricted ansatz expressivity, and reliance on domain-specific structure. We present SpinGQE, an extension of the Generative Quantum Eigensolver (GQE) framework to spin Hamiltonians. Our approach reframes circuit design as a generative modeling task. We employ a transformer-based decoder to learn distributions over quantum circuits that produce low-energy states. Training is guided by a weighted mean-squared error loss between model logits and circuit energies evaluated at each gate subsequence. We validate our method on the four-qubit Heisenberg model, demonstrating successfulconvergencetonear-groundstates. Throughsystematichyperparameterexploration, we identify optimal configurations: smaller model architectures (12 layers, 8 attention heads), longer sequence lengths (12 gates), and carefully chosen operator pools yield the most reliable convergence. Our results show that generative approaches can effectively navigate complex energy landscapes without relying on problem-specific symmetries or structure. This provides a scalable alternative to traditional variational methods for general quantum systems. An open-source implementation is available at https://github.com/Mindbeam-AI/SpinGQE.

We study word-level semantic alignment across four historical stages of Ancient Egyptian. These stages differ in script and orthography, and parallel data are scarce. We jointly train a compact encoder-decoder model with a shared byte-level tokenizer on all four stages, combining masked language modeling (MLM), translation language modeling (TLM), sequence-to-sequence translation, and part-of-speech tagging under a task-aware loss with fixed weights and uncertainty-based scaling. To reduce surface divergence we add Latin transliteration and IPA reconstruction as auxiliary views. We integrate these views through KL-based consistency and through embedding-level fusion. We evaluate alignment quality using pairwise metrics, specifically ROC-AUC and triplet accuracy, on curated Egyptian-English and intra-Egyptian cognate datasets. Translation yields the strongest gains. IPA with KL consistency improves cross-branch alignment, while early fusion demonstrates limited efficacy. Although the overall alignment remains limited, the findings provide a reproducible baseline and practical guidance for modeling historical languages under real constraints. They also show how normalization and task design shape what counts as alignment in typologically distant settings.

This paper describes the system submitted to the SOMD 2026 Shared Task for Cross-Document Coreference Resolution (CDCR) of software mentions. Our approach addresses the challenge of identifying and clustering inconsistent software mentions across scientific corpora. We propose a hybrid framework that combines dense semantic embeddings from a pre-trained Sentence-BERT model, Knowledge Base (KB) lookup strategy built from training-set cluster centroids using FAISS for efficient retrieval, and HDBSCAN density-based clustering for mentions that cannot be confidently assigned to existing clusters. Surface-form normalization and abbreviation resolution are applied to improve canonical name matching. The same core pipeline is applied to Subtasks 1 and 2. To address the large scale settings of Subtask 3, the pipeline was adapted by utilising a blocking strategy based on entity types and canonicalized surface forms. Our system achieved CoNLL F1 scores of 0.98, 0.98, and 0.96 on Subtasks 1, 2, and 3 respectively.

Federated Learning (FL) of Large Language Models (LLMs) in multilingual environments presents significant challenges stemming from heterogeneous language distributions across clients and disparities in language resource availability. To address these challenges, we extended the FederatedScope-LLM framework to support multilingual instruction-tuning experiments with LLMs. We also introduced a novel client-specific early stopping mechanism, Local Dynamic Early Stopping (LDES-FL), which allows clients to pause and resume local training based on client-side validation performance, enhancing training efficiency and sustainability. Through a series of experiments, we studied how client language composition - from fully monolingual to increasingly multilingual clients - affects multilingual quality, fairness and training cost. Monolingual local fine-tuning remains the most effective for single-language specialization, whereas federated training is better suited to learning a single balanced multilingual model. In FL, increasing within-client multilinguality leads to stronger and fairer global models, narrows the gap to centralized multilingual fine-tuning, and yields the largest gains for lower-resource languages, albeit at the cost of more optimization steps. Overall, our results identify client language composition as a key design variable in multilingual FL, shaping performance, fairness and efficiency.

Stance detection is nearly always formulated as classifying text into Favor, Against, or Neutral. This convention was inherited from debate analysis and has been applied without modification to social media since SemEval-2016. However, attitudes toward complex targets are not unitary. A person can accept climate science while opposing carbon taxes, expressing support on one dimension and opposition on another. When annotators must compress such multi-dimensional attitudes into a single label, different annotators may weight different dimensions, producing disagreement that reflects different compression choices rather than confusion. We call this the projection problem. We conduct an annotation study across five targets from three stance benchmarks (SemEval-2016, P-Stance, COVID-19-Stance), with the same three annotators labeling all targets. For each target, annotators assign both a standard stance label and per-dimension judgments along target-specific dimensions discovered through bottom-up analysis, using the same number of categories for both. Across all fifteen target–dimension pairs, dimensional agreement consistently exceeds label agreement. The gap appears to scale with target complexity: modest for a single-entity target like Joe Biden (AC1: 0.87 vs. 0.95), but large for a multi-faceted policy target like school closures (AC1: 0.21 vs. 0.71).

In Natural Language Processing (NLP), variation is typically seen as noise and "normalised away" before processing, even though it is an integral part of language. Conversely, studying language variation in social contexts is central to sociolinguistics. We present a framework to combine the sociolinguistic dimension of language with the technical dimension of NLP. We argue that by embracing sociolinguistics, variation can actively be included in a research setup, in turn informing the NLP side. To illustrate this, we provide a case study on Luxembourgish, an evolving language featuring a large amount of orthographic variation, demonstrating how NLP performance is impacted. The results show large discrepancies in the performance of models tested and fine-tuned on data with a large amount of orthographic variation in comparison to data closer to the (orthographic) standard. Furthermore, we provide a possible solution to improve the performance by including variation in the fine-tuning process. This case study highlights the importance of including variation in the research setup, as models are currently not robust to occurring variation. Our framework facilitates the inclusion of variation in the thought-process while also being grounded in the theoretical framework of sociolinguistics.

Antonyms, or opposites, are sometimes defined as word pairs that have all of the same contextually relevant properties but one. Seeing how transformer models seem to encode concepts as directions, this begs the question if one can detect “antonymity” in the geometry of the embedding vectors of word pairs, especially based on their difference vectors. Such geometrical studies are then naturally contrasted by comparing antonymic pairs to their opposites; synonyms. This paper started as an exploratory project on the complexity of the systems needed to detect the geometry of the embedding vectors of antonymic word pairs. What we now report is a curious “swirl” that appears across embedding models in a somewhat specific projection configuration.

Conversational artificial intelligence has the potential to assist users in preliminary medical consultations, particularly in settings where access to healthcare professionals is limited. However, many existing medical dialogue systems operate in a single-turn question–answering paradigm or rely on template-based datasets, limiting conversational realism and multilingual applicability. In this work, we introduce MedAidDialog, a multilingual multi-turn medical dialogue dataset designed to simulate realistic physician–patient consultations. The dataset extends the MDDial corpus by generating synthetic consultations using large language models and further expands them into a parallel multilingual corpus covering seven languages: English, Hindi, Telugu, Tamil, Bengali, Marathi, and Arabic. Building on this dataset, we develop MedAidLM, a conversational medical model trained using parameter-efficient fine-tuning on quantized small language models, enabling deployment without high-end computational infrastructure. Our framework additionally incorporates optional patient pre-context information (e.g., age, gender, allergies) to personalize the consultation process. Experimental results demonstrate that the proposed system can effectively perform symptom elicitation through multi-turn dialogue and generate diagnostic recommendations. We further conduct medical expert evaluation to assess the plausibility and coherence of the generated consultations.

During training, Large Language Models (LLMs) learn social regularities that can lead to gender bias in downstream applications. Most mitigation efforts focus on reducing bias in generated outputs, typically evaluated on structured benchmarks, which raises two concerns: output-level evaluation does not reveal whether alignment modifies the model's underlying representations, and structured benchmarks may not reflect realistic usage scenarios. We propose a unified framework to jointly analyze intrinsic and extrinsic gender bias in LLMs using identical neutral prompts, enabling direct comparison between gender-related information encoded in internal representations and bias expressed in generated outputs. Contrary to prior work reporting weak or inconsistent correlations, we find a consistent association between latent gender information and expressed bias when measured under the unified protocol. We further examine the effect of alignment through supervised fine-tuning aimed at reducing gender bias. Our results suggest that while the latter indeed reduces expressed bias, measurable gender-related associations are still present in internal representations, and can be reactivated under adversarial prompting. Finally, we consider two realistic settings and show that debiasing effects observed on structured benchmarks do not necessarily generalize, e.g., to the case of story generation.

RLHF-aligned language models exhibit response homogenization: on TruthfulQA (n=790), 40-79% of questions produce a single semantic cluster across 10 i.i.d. samples. On affected questions, sampling-based uncertainty methods have zero discriminative power (AUROC=0.500), while free token entropy retains signal (0.603). This alignment tax is task-dependent: on GSM8K (n=500), token entropy achieves 0.724 (Cohen's d=0.81). A base-vs-instruct ablation confirms the causal role of alignment: the base model shows 1.0% single-cluster rate vs. 28.5% for the instruct model (p < 10^{-6}). A training stage ablation (Base 0.0% -> SFT 1.5% -> DPO 4.0% SCR) localizes the cause to DPO, not SFT. Cross-family replication on four model families reveals alignment tax severity varies by family and scale. We validate across 22 experiments, 5 benchmarks, 4 model families, and 3 model scales (3B-14B), with Jaccard, embedding, and NLI-based baselines at three DeBERTa scales (all ~0.51 AUROC). Cross-embedder validation with two independent embedding families rules out coupling bias. Cross-dataset validation on WebQuestions (58.0% SCR) confirms generalization beyond TruthfulQA. The central finding – response homogenization – is implementation-independent and label-free. Motivated by this diagnosis, we explore a cheapest-first cascade (UCBD) over orthogonal uncertainty signals. Selective prediction raises GSM8K accuracy from 84.4% to 93.2% at 50% coverage; weakly dependent boundaries (|r| <= 0.12) enable 57% cost savings.

Benchmarks like MMLU suggest flagship language models approach factuality saturation above 90%. LLMpedia shows this picture is incomplete. We materialize ${\sim}$1.3M encyclopedia articles entirely from parametric memory across three model families, then audit every claim against Wikipedia and curated web evidence. For gpt-5-mini, the verifiable true rate is 68.4% on Wikipedia-covered subjects - more than 21\,pp below MMLU - and the gap is driven by unverifiability (30.5%), not refutation (1.2%). Beyond Wikipedia, frontier articles audited against curated web evidence reach 57.6%; Wikipedia covers only 56.7% of model-surfaced subjects, and three model families overlap in just 7.3% of subject choices. In a retrieval-trap benchmark inspired by prior analysis of Grokipedia, LLMpedia is more factual at roughly half the textual similarity to Wikipedia. Every prompt, article, and verdict is released. Data, code, interface: https://llmpedia.net.

Knowledge Tracing (KT) is a critical technique for modeling student knowledge to support personalized learning. However, most KT systems focus on binary correctness prediction and cannot diagnose the underlying conceptual misunderstandings that lead to errors. Such fine-grained diagnostic feedback is essential for designing targeted instruction and effective remediation. In this work, we introduce the task of concept-level deficiency prediction, which extends traditional KT by identifying the specific concepts a student is likely to struggle with on future problems. We present ConceptKT, a dataset annotated with labels that capture both the concepts required to solve each question and the missing concepts underlying incorrect responses. We investigate in-context learning approaches to KT and evaluate the diagnostic capabilities of various Large Language Models (LLMs) and Large Reasoning Models (LRMs). Different strategies for selecting informative historical records are explored. Experimental results demonstrate that selecting response histories based on conceptual alignment and semantic similarity leads to improved performance on both correctness prediction and concept-level deficiency identification.

Tool-use capabilities are vital for Large Language Models (LLMs) in finance, a domain characterized by massive investment targets and data-intensive inquiries. However, existing data synthesis methods typically rely on a reverse synthesis paradigm, generating user queries from pre-sampled tools. This approach inevitably introduces artificial explicitness, yielding queries that fail to capture the implicit, event-driven nature of real-world needs. Moreover, its reliance on static tool sets overlooks the dynamic retrieval process required to navigate massive tool spaces. To address these challenges, we introduce FinToolSyn, a forward synthesis framework designed to generate high-quality financial dialogues. Progressing from persona instruction and atomic tool synthesis to dynamic retrieval dialogue generation, our pipeline constructs a repository of 43,066 tools and synthesizes over 148k dialogue instances, incorporating dynamic retrieval to emulate the noisy candidate sets typical of massive tool spaces. We also establish a dedicated benchmark to evaluate tool-calling capabilities in realistic financial scenarios. Extensive experiments demonstrate that models trained on FinToolSyn achieve a 21.06% improvement, providing a robust foundation for tool learning in financial scenarios.

Standard LoRA fine-tuning of Mixture-of-Experts (MoE) models applies adapters to every expert, yet our profiling shows that per-layer expert routing is highly skewed: a small subset of experts handles most tokens in each layer, while many others are rarely activated ("cold"). We propose MoE-Sieve, a simple routing-guided framework for LoRA fine-tuning, and pair it with a systematic profiling study of expert routing across architectures and tasks. The method is simple: profile routing counts on a small calibration set, select the top-k most-routed experts per layer, and apply LoRA only to those experts. Across two architecturally distinct MoE models and three diverse tasks, tuning only the top 25% routed experts per layer remains competitive with full LoRA, with mean differences within +/-1 percentage point across all conditions. This reduces LoRA trainable parameters by 70-73%, adapter checkpoint size by 71-73%, and wall-clock training time by up to 50%. We also observe a non-monotonic relationship between expert count and seed-to-seed variance, consistent with the hypothesis that adapting cold experts can introduce gradient noise without improving accuracy. Further ablations show that random expert selection at matched budget is about 2.5 percentage points worse, indicating that the routing signal matters, while greedy per-layer budget optimization does not improve over uniform top-k.