Papers

Repeated sampling can improve LLM accuracy, and quantifying the gains from additional calls is essential for allocating test-time compute. We study binary decisions, a fundamental setting where repeated answers to the same question are aggregated by majority vote. We show that two independently sampled responses per example in a validation set with known answers constrain the latent distribution of example-level success probabilities to a class consistent with the paired outcomes. Optimizing over this class yields sharp accuracy and gain bounds at every finite voting budget. A shared large-sample confidence region accounts for validation uncertainty. We also obtain sharp infinite-vote bounds and moment-matched forecasts of the vote-accuracy curve. On QNLI and QQP, our method successfully distinguishes settings with voting gains above a target margin from those with little room for improvement.

Retrieval-augmented generation (RAG) has proven effective for knowledge-intensive tasks, but is widely believed to offer limited benefit for reasoning-intensive problems such as math and code generation. We challenge this assumption by showing that the limitation lies not in RAG itself, but in the choice of corpus. Instead of retrieving documents, we propose retrieving thinking traces, i.e., intermediate thinking trajectories generated during problem solving attempts. We show that thinking traces are already a strong retrieval source, and further introduce T3, an offline method that transforms them into structured, retrieval-friendly representations, to improve usability. Using these traces as a corpus, a simple retrieve-then-generate pipeline consistently improves reasoning performance across strong models and benchmarks such as AIME 2025–2026, LiveCodeBench, and GPQA-Diamond, outperforming both non-RAG baselines and retrieval over standard web corpora. For instance, on AIME 2025-2026, RAG with traces generated by Gemini-2-thinking achieves relative gains of +56.3%, +8.6%, and +7.6% for Gemini-2.5-Flash, GPT-OSS-120B, and GPT-5, respectively, even though these are more recent models. Overall, our results suggest that thinking traces are an effective retrieval corpus for reasoning tasks, and transforming them into structured, compact, or diagnostic representations unlocks even stronger gains. Code available at https://github.com/Narabzad/t3.

In single-stream autoregressive interfaces, the same tokens both update the model state and constitute an irreversible public commitment. This coupling creates a silence tax: additional deliberation postpones the first task-relevant content, while naive early streaming risks premature commitments that bias subsequent generations. We introduce Side-by-Side (SxS) Interleaved Reasoning, which makes disclosure timing a controllable decision within standard autoregressive generation. SxS interleaves partial disclosures with continued private reasoning in the same context, but releases content only when it is supported by the reasoning so far. To learn such pacing without incentivizing filler, we construct entailment-aligned interleaved trajectories by matching answer prefixes to supporting reasoning prefixes, then train with SFT to acquire the dual-action semantics and RL to recover reasoning performance under the new format. Across two Qwen3 architectures/scales (MoE Qwen3-30B-A3B, dense Qwen3-4B) and both in-domain (AIME25) and out-of-domain (GPQA-Diamond) benchmarks, SxS improves accuracy–content-latency Pareto trade-offs under token-level proxies such as inter-update waiting.

De-identification of clinical text is a prerequisite for the secondary use of electronic health records. Existing public benchmarks such as the i2b2 2006 and 2014 corpora are over a decade old and lack the semantic and demographic diversity of modern clinical narratives. Large Language Models (LLMs) reach state-of-the-art zero-shot extraction, but their use at enterprise scale is limited by computational cost and by hospital data governance that restricts sending Protected Health Information (PHI) to cloud APIs. We introduce SHIELD (Synthetic Human-annotated Identifier-replaced Entries for Learning and De-identification), a diverse clinical note dataset of 1,381 notes with 10,229 gold-standard PHI spans across 9 categories, built with set-cover diversity sampling across demographic and document-type strata and human-in-the-loop adjudication. We evaluate four LLMs (two proprietary, two open-weight) to establish a performance ceiling on SHIELD, then show that a teacher-student distillation framework transfers these capabilities into locally deployable Small Language Models. Our best distilled model reaches micro-averaged span-level precision of 0.89 and recall of 0.88 while running on standard workstation hardware. It trails its cloud teacher on per-category recall (0.90 vs. 0.81 macro-averaged) but remains competitive given its lower cost and on-premise deployability. Cross-dataset evaluation shows that diversity-trained models generalize well on universal structured PHI categories, while institution-specific entities remain hard to transfer in both directions, which suggests pairing broad-coverage models with specialized models for high-volume, semi-structured note types. We publicly release the SHIELD dataset and the distilled DeBERTa v3 model to provide an accurate, cost-effective de-identification pipeline deployable entirely behind institutional firewalls.

Cross-lingual topic modeling aims to discover shared semantic structures across languages, yet existing models depend on sparse bilingual resources and often yield incoherent or weakly aligned topics. Recent LLM-based refinements improve interpretability but are costly, document-level, and prone to hallucination, with prior white-box approaches requiring inaccessible token probabilities. We propose LLM-XTM, a framework that integrates LLM-guided topic refinement with self-consistency uncertainty quantification, enabling black-box, stable, and scalable enhancement of cross-lingual topic models. Experiments on multilingual corpora show that LLM-XTM achieves superior topic coherence and alignment while reducing reliance on bilingual dictionaries and expensive LLM calls.

Large language models often fail at simple counting tasks, even when items to count are in the prompt. We investigate whether this failure occurs because transformers do not represent counts internally, or because they cannot convert representations to the correct output tokens. Across three model families: Pythia, Qwen3, and Mistral, ranging from 0.4B to 14B parameters, we find evidence for the second explanation. Linear probes recover the correct count from intermediate layers with $R^2>0.99$, showing that the information is present. However, the internal directions that encode counts are nearly orthogonal to digit-token output-head rows ($|\cos| \leq 0.032$). In other words, the model stores the count in a form that the digit logits do not naturally read out. We localize this failure with two interventions. Updating only the digit rows of the output head (36,864 parameters) substantially improves constrained digit prediction (60.7–100.0% on four tasks), but it does not fix unconstrained generation (0%); we do not claim that digit-row repair fixes open-ended text. By contrast, small LoRA on attention Q/V (7.67M parameters) improves upstream routing and achieves 83.1%$\pm$7.2% in true greedy autoregressive generation (deployable fix). Logit-lens at layer 35 (entity counting; correct-digit rank): (i) median over 3 seeds drops from order-$10^4$ to 1; (ii) seed 42 shows $54{,}332 \to 838$ (median top-1 while one seed stays far below). Norm, logit-lens, and cross-task analyses generalize the bottleneck to counting, addition, and list length; nulls on MMLU and GSM8K and limited DROP transfer. These results identify counting failure as a geometric readout bottleneck, not an internal-representation failure: the model knows the count but the output pathway is misaligned with tokens needed to express it.

Movie scripts pose a fundamental challenge for automatic summarization due to their non-linear, cross-cut narrative structure, which makes surface-level saliency methods ineffective at preserving core story progression. To address this, we introduce S^2tory (Story Spine Distillation), a narratology-grounded framework that leverages character development trajectories to identify plot nuclei, the essential events that drive the narrative forward, while filtering out peripheral satellite events that merely enrich atmosphere or emotion. Our Narrative Expert Agent (NEAgent) performs theory-constrained reasoning, whose distilled knowledge conditions a small model to identify plot nuclei. Another model then uses these plot nuclei to generate the summary. Experiments on the MovieSum dataset demonstrate state-of-the-art semantic fidelity at approximately 3.5x compression, and zero-shot evaluation on BookSum confirms strong out-of-domain generalization. Human evaluation further validates that narratological theory provides an indispensable foundation for modeling complex, non-linear narratives.

Project-based learning improves student engagement and learning outcomes, yet allocating students to appropriately challenging projects while forming cognitively diverse teams remains difficult at scale. Traditional allocation methods (manual spreadsheets, preference surveys) can't construct the cognitively diverse teams that that collaborate cognitively. This mismatch perpetuates equity issues: high-performing students self-select visible projects while under-represented students face reduced access to opportunity. We propose TeamUp, a lightweight, embedding-based team-forming system designed to improve learning outcomes and equity in large-scale project-based courses. TeamUp uses semantic embeddings from pretrained language models to match students to projects aligned with their skill level. The system employs a hybrid ranking algorithm combining cosine similarity with pedagogical constraints (difficulty alignment, domain preferences, and demand balancing) to generate personalised and transparent recommendations. Beyond individual matching, TeamUp constructs cognitively diverse teams by modelling skill complementarity through embedding variance, ensuring teams possess well-distributed capabilities rather than homogeneous strengths. We evaluated TeamUp through a virtual experiment using 250 student profiles and 60 project descriptions. Results show: (1) substantially higher match quality (mean cosine similarity of 0.74 vs. 0.43); (2) better difficulty alignment (83% placed within one level vs. 34%); (3) more diverse teams (82% covering three or more technical areas vs. 41%); and (4) sub-second recommendation latency at operational costs under $0.10 per student.

Large language models (LLMs) are increasingly used by end users, yet existing personalization methods relying on static profiles or text-only signals fail to capture query-specific expertise variation. We present ExPerT, a query-wise personalization framework that adapts LLM responses to users' query domain expertise by combining semantic and behavioral cues. ExPerT consists of two key components: (i) a semantic-behavioral expertise inference module that jointly interprets query text and keystroke dynamics via in-context LLM prompting, and (ii) an expertise-conditioned response generation that adapts the level of detail, terminology, and conceptual complexity. Our user study with 40 participants and 1270 queries demonstrated that ExPerT reduced expertise inference error by 65.7% compared to the strongest baseline (MAE = 0.398 vs. 1.162) and improved response satisfaction by 17.52% (from 3.71 to 4.36) on a 5-point Likert scale.

Existing prompt compression methods treat text as flat token sequences, failing to capture the distributed nature of important information, which is often spread across multiple locations and connected through both local syntactic dependencies and global semantic relations. Such relational structure is naturally represented as a graph, where tokens or sentences become nodes and their dependencies become edges. To this end, we propose RAGP, which formulates prompt compression as Redundancy-Aware Graph Pruning on a multiplex graph that jointly models fine-grained attention-based dependencies and coarse-grained semantic relations. To efficiently identify non-redundant nodes in this heterogeneous structure (dense local subgraphs and sparse global connections), we employ Levy walks whose heavy-tailed step distribution naturally balances local exploitation with global exploration. Experiments on LongBench show that RAGP achieves an average score of 49.3 under a 4x compression ratio, outperforming existing LLM-based compression methods, such as LongLLMLingua, which attains 48.8 at a 3x compression ratio. Besides, RAGP also surpasses state-of-the-art vision-based text compression paradigms on multiple tasks.

Sign language communication systems, that integrate emotional expression remain underexplored, particularly for low-resource languages. This pilot study presents NEST-V1 (Nepali Emotion and Speech Transformer - Version 1), a proof-of-concept multimodal framework that demonstrates the feasibility of generating emotion-conditioned Nepali Sign Language avatars from spoken input. As a preliminary investigation, we focus on four common Nepali words ("thank you", "hello", "house", "me") across three emotional states (happy, neutral, sad) to validate our core technical approach. Our lightweight architecture employs a shared acoustic encoder for simultaneous Automatic Speech Recognition and emotion classification, achieving 81.1% ASR accuracy and 79.21% emotion recognition accuracy on a dataset of 600 labeled audio samples from 50 speakers. The system demonstrates 37% parameter efficiency compared to separate model architectures while maintaining a lightweight footprint with only 22.1M parameters suitable for edge deployment. This pilot work establishes the technical foundation for emotion-aware sign language translation in low-resource settings and provides a scalable framework for future expansion to larger vocabularies and more diverse emotional expressions. Our preliminary results indicate the viability of real-time, emotionally expressive sign language communication systems for the hearing-impaired community, with clear pathways for enhancement in subsequent development phases.

Speech-based large language models are typically constrained to spoken replies, which limits their user-facing outputs to what can be verbalized and suppresses text-native capabilities such as code generation, structured analysis, and multi-step reasoning in realtime interaction, for tasks that require persistent, structured, and inspectable intermediate outputs. Existing work improves spoken reasoning or full-duplex turn-taking, but still treats text as a hidden intermediate state or a subordinate modality rather than a first-class output channel. We propose Listen-Write-Speak (LWS), a text-first tri-channel paradigm in which a single autoregressive LLM continuously listens to user audio, writes visible free-form text as its primary output, and speaks a realtime oral response in parallel under a shared causal attention context. This behavior is implemented entirely through a Token Schema, requiring no architectural modifications, and learned via a two-stage data pipeline that synthesizes per-second cognitive annotations consistent with the revealed input timeline. Empirically, LWS demonstrates strong full-duplex interaction on Full-Duplex-Bench, reaches 4.72 on VoiceBench AlpacaEval, achieves 92.6% writing-speaking consistency, and consistently outperforms its internal ablations on URO-Bench. These results suggest that visible writing can serve as a first-class output channel for speech interaction without sacrificing realtime responsiveness. The code and dataset are available on the project page: https://royalzhang.com/project/lws-page/.

Text-in-image editing has become a key capability for visual content creation, yet existing benchmarks remain overwhelmingly English-centric and often conflate visual plausibility with semantic correctness. We introduce MULTITEXTEDIT, a controlled benchmark of 3,600 instances spanning 12 typologically diverse languages, 5 visual domains, and 7 editing operations. Language variants of each instance share a common visual base and are paired with a human-edited reference and region masks, isolating the language variable for cross-lingual comparison. To capture script-level errors that coarse text-matching metrics miss, such as missing diacritics, reversed RTL order, and mixed-script renderings, we introduce a language fidelity (LSF) metric scored by a two-stage LVM protocol that first traces the edited target text and then judges it in isolation, reaching a quadratic-weighted ąppa} of 0.76 against native-speaker annotators. Evaluating 12 open-source and proprietary systems with LSF alongside standard semantic and mask-aware pixel metrics, we find pronounced cross-lingual degradation for every model, largest on Hebrew and Arabic and smallest on Dutch and Spanish, and concentrated in text accuracy and script fidelity rather than in coarse structural dimensions. We also uncover a pervasive semantic and pixel mismatch, where outputs preserve global layout and background fidelity yet distort script-specific forms.

Retrieval-augmented generation (RAG) remains brittle on multi-hop questions in realistic deployment settings, where retrieved evidence may be noisy or redundant and only limited context can be passed to the generator. Existing controllers address parts of this problem, but typically either expand context additively, select from a fixed top-k set, or optimize relevance without explicitly repairing missing bridge facts. We propose AdaGATE, a training-free evidence controller for multi-hop RAG that frames evidence selection as a token-constrained repair problem. AdaGATE combines entity centric gap tracking, targeted micro-query generation, and a utility based selection mechanism that balances gap coverage, corroboration, novelty, redundancy, and direct question relevance. We evaluate AdaGATE on HotpotQA under clean, redundancy, and noise injected retrieval conditions. Across all three settings, AdaGATE achieves the best evidence F1 among the compared controllers, reaching 62.3% on clean data and 71.2% under redundancy injection, while using 2.6x fewer input tokens than Adaptive-k. These results suggest that explicit gap-aware repair, combined with token-efficient evidence selection, improves robustness in multi-hop RAG under imperfect retrieval. Our code and evaluation pipeline are available at https://github.com/eliguo/AdaGATE.

Production agent frameworks (OpenAI Function Calling, Anthropic Tool Use, MCP) transmit tool schemas as JSON, a format designed for machine parsing, not for interpretation by language models. For small models (4B-14B), this protocol mismatch accounts for the majority of tool-use failure at production catalog sizes. We present TSCG, a deterministic tool-schema compiler that resolves this mismatch at the API boundary, converting JSON schemas into token-efficient structured text without model access, fine-tuning, or runtime search. TSCG combines eight composable operators with a formal compression bound (>=51% on well-formed schemas). On TSCG-Agentic-Bench (about 19,000 calls, 12 models, 5 scenarios), TSCG restores Phi-4 14B from 0% to 84.4% accuracy at 20 tools (90.3% at 50 tools) and achieves 108-181% accuracy-retained ratio across three models on BFCL. Format-versus-compression decomposition (R^2=0.88 -> 0.03) establishes representation change as the dominant mechanism. Per-operator isolation across three frontier models reveals three distinct operator-response profiles: operator-hungry (Opus 4.7), operator-sensitive (GPT-5.2), and operator-robust (Sonnet 4), providing per-model deployment guidance. Scaling experiments show accuracy advantages persisting on heavy production MCP schemas (+5.0 pp at about 10,500 input tokens) despite saturation on light synthetic catalogs, with 52-57% token savings throughout. The synthetic benchmark generalizes to real MCP schemas within 0.1 accuracy points. TSCG ships as a 1,200-line zero-dependency TypeScript package.

Adapting a pretrained language model to a new task often hurts the general capabilities it already had, a problem known as catastrophic forgetting. Sparse Memory Finetuning (SMF) tries to avoid this by adding key-value memory layers to the model and, on each training step, updating only the small set of memory rows that the current batch reads most heavily. We re-implement SMF on Qwen-2.5-0.5B-Instruct and compare it with LoRA and full finetuning on MedMCQA, a 4-choice medical exam task, using WikiText perplexity and TriviaQA accuracy as forgetting probes. SMF improves MedMCQA by 2.5 percentage points while keeping both forgetting probes within roughly 1 point of the base model, whereas LoRA and full finetuning achieve larger gains but with clear drift on both. We also compare two row-selection rules (KL-divergence and TF-IDF), which balance the two forgetting metrics differently.

As large language model (LLM)-powered agents are increasingly deployed to perform complex, real-world tasks, they face a growing class of attacks that exploit extended user-agent-environment interactions to pursue malicious objectives improbable in single-turn settings. Such long-horizon threats pose significant risks to the safe deployment of LLM agents in critical domains. In this paper, we present ShadowMem, a novel defensive framework designed to counter a wide range of long-horizon threats. Inspired by the "shadow stack" abstraction in systems security, ShadowMem maintains a dedicated, safety-focused agentic memory that distills and retains safety-critical context across the agent's full execution trajectory, leveraging this shadow memory to proactively assess the risk of pending actions prior to their execution. Extensive evaluation demonstrates that ShadowMem substantially outperforms existing defenses across diverse long-horizon threats in detection accuracy, achieves early-stage detection for the majority of attacks, and introduces only negligible overhead to agent utility. To our best knowledge, ShadowMem represents the first framework to detect and mitigate long-horizon threats using an agentic memory approach, establishing a new paradigm for this critical challenge and opening promising directions for future research. The artifacts are available at https://github.com/ZJUWYH/ShadowMem

Modeling latent clinical constructs from unconstrained clinical interactions is a unique challenge in affective computing. We present ADAPTS (Agentic Decomposition for Automated Protocol-agnostic Tracking of Symptoms), a framework for automated rating of depression and anxiety severity using a mixture-of-agents LLM architecture. This approach decomposes long-form clinical interviews into symptom-specific reasoning tasks, producing auditable justifications while preserving temporal and speaker alignment. Generalization was evaluated across two independent datasets ($N=204$) with distinct interview structures. On high-discrepancy interviews, automated ratings approximated expert benchmarks ($\text{absolute error}=22$) more closely than original human ratings ($\text{absolute error}=26$). Implementing an “extended” protocol that incorporates qualitative clinical conventions significantly stabilized ratings, with absolute agreement reaching $\text{ICC(2,1)} = 0.877$. These findings suggest that the ADAPTS framework enables promising evaluations of psychiatric severity. While the current implementation is purely text-based, the underlying architecture is readily extensible to multimodal inputs, including acoustic and visual features. By approximating expert-level precision in a protocol-agnostic manner, this framework provides a foundation for objective and scalable psychiatric assessment, especially in resource-limited settings.

A reliable language model should be able to signal, prior to generation, when a query falls outside its knowledge. We investigate whether representation geometry can provide such a pre-generation signal by measuring the deviation of hidden states from an answerable reference set, requiring no labeled failure data and no access to model outputs. Across three instruction-tuned models (Llama 3.1-8B, Qwen 2.5-7B, and Mistral-7B-Instruct) and three prompt forms (Math, Fact, Code), we find that geometry primarily encodes task form. Within mathematical prompts, unanswerable inputs consistently deviate from the answerable centroid, yielding strong separation (ROC-AUC 0.78-0.84). This single-pass pre-generation signal outperforms a simple refusal baseline and compares favorably to self-consistency. It also captures cases where models do not explicitly refuse. In contrast, no reliable geometric signal emerges for factual prompts, indicating that the effect is form-conditional rather than universal. Code prompts show large effect sizes with higher variance, suggesting partial generalization beyond mathematical form. A layer-wise analysis reveals that the signal arises in early layers and gradually attenuates toward the output. These results suggest that answerability-related geometry is established before the final stages of generation. Together, these findings indicate that geometric deviation can serve as a lightweight pre-generation signal that is reliable in structured domains with formal answerability constraints, with clear boundaries on where it generalizes.

Static benchmarks measure a model frozen at training time. Real systems face distribution shift: new categories, paraphrased queries, drift: and must recover online via user corrections. No existing benchmark measures recovery speed under correction streams. We introduce OCRR (Online Correction Recovery Rate): a benchmark that streams a corpus through a classification system, applies oracle or stochastic corrections to wrong predictions, and reports two curves: novel-class accuracy and original-distribution accuracy versus correction count. We evaluate the substrate alongside nine baseline algorithms from five families plus seven bounded-storage variants of the substrate for the Pareto sweep, including standard online-learning baselines (river), continual-learning methods (EWC, A-GEM, LwF), retrieval/parametric hybrids (kNN-LM), parameter-efficient fine-tuning of a 1.5 B-parameter encoder (LoRA on DeBERTa-v3-large), and a hash-chained append-only substrate (Substrate). On Banking77 and CLINC150, under oracle and sparse correction policies, the substrate is the only system that simultaneously recovers novel-class accuracy (88.7 +/- 2.9 %) and retains original-distribution accuracy (95.4 +/- 0.8 %) beating the next-best published continual-learning baseline by 32.6 percentage points at equal memory budget, and beating LoRA-on-DeBERTa-v3-large by 84.6 percentage points on retention. We further find that classification accuracy remains stable at 99 % even as approximate-nearest-neighbour recall@5 degrades from 0.69 to 0.23 across 10 k to 10 M corpus scales, suggesting the substrate's margin-band majority vote is robust to retrieval imperfection in a way that pure top-k recall metrics do not predict. Code and data are available at https://github.com/adriangrassi/ocrr-benchmark.

Earnings calls are a key source of financial information about public companies. However, extracting information from these calls is difficult. Unlike the templatic filings required by the U.S. Securities and Exchange Commission (SEC) to report a company's financial situation, earnings conference calls have no built-in labels, are unstructured, and feature conversational language. We explore this challenging domain by assessing the information captured by models trained on SEC filings and in-context learning methods. To establish a baseline, we first evaluate the generalization capabilities of SEC-trained models across established SEC datasets. To support our investigation, we introduce three novel benchmarks: (1) SEC Filings Benchmark (SECB), (2) Earnings Calls Benchmark (ECB), and ECB-A, a subset with 2,460 expert annotation groups to support our qualitative analysis. We find that encoder-based models struggle with the domain shift. Finally, we propose a system utilizing LLMs to perform open-ended extraction from unstructured call transcripts, verified by human evaluation (79.7% precision), providing a baseline for this valuable domain through the consistent tracking of emergent KPIs.

Browsing-enabled LLM assistants can fetch webpages and answer contact-seeking queries, creating a practical channel for scraping contact-style personally identifiable information (PII) from public pages. Many prior defenses are deployed at the model, service, or agent layer rather than at the webpage itself, leaving ordinary page owners with limited deployable options. We present PIIGuard, a webpage-level defense that repurposes indirect prompt injection as a protective mechanism: the page owner embeds optimized hidden HTML fragments that steer the model away from verbatim or reconstructible disclosure of contact PII. PIIGuard searches over fragment text and insertion position using rule-based leakage scoring, evolutionary mutation, and final judge-based recoverability assessment. In direct-HTML evaluation on three target models (GPT-5.4-nano, Claude-haiku-4.5, and DeepSeek-chat(latest v3.2)), PIIGuard achieves at least 97.0% defense success rate under both rule-based and judge-based leakage evaluation, often reaching 100.0%, while preserving benign same-page QA utility. We further evaluate two harder settings: public-URL browsing and attacker-side LLM sanitization of fetched webpage. These results show that page-side defensive fragments can remain effective in deployment for some model-position pairs, but robustness varies substantially across browsing interfaces and sanitizer prompts. Overall, PIIGuard demonstrates that page owners can use page-side fragments as a practical mitigation for web-grounded PII leakage.

Social-educational robots designed for socially interactive pedagogical support, such as the Robot Study Companion (RSC), rely on responsive, privacy-preserving interaction despite severely limited compute. However, there is a gap in systematic benchmarking of language models for edge computing in pedagogical applications. This paper benchmarks 25 open-source language models for local deployment on edge hardware. We evaluate each model across three dimensions: inference efficiency (tokens per second, energy consumption), general knowledge (a six-category MMLU subset), and teaching effectiveness (LLM-rated pedagogical quality), validated against five independent human raters using the Raspberry Pi(RPi)4 as the primary platform, with additional comparisons on the RPi5 and a laptop GPU. Results reveal pronounced trade-offs: throughput and energy efficiency vary by over an order of magnitude across models, MMLU accuracy ranges from near-random to 57.2%, and teaching effectiveness does not correlate monotonically with either metric. Among the evaluated models, Granite4 Tiny Hybrid (7B) achieves a strong overall balance, reaching 2.5 tokens per second, 0.90 tokens per joule, and 54.6% MMLU accuracy; high MMLU accuracy does not appear necessary for strong teaching scores. Human validation on four representative models preserved the automated rank ordering (Pearson r = 0.967, n = 4). Based on these findings, we propose a three-tier local inference architecture for the RSC that balances responsiveness and accuracy on resource-constrained hardware.

Semi-structured information extraction (IE) from OCR-derived clinical reports is crucial for efficiently reconstructing patients' longitudinal medical histories. In practice, this scenario commonly involves three tasks: (i) field-header (key) discovery, (ii) key-conditioned question answering (QA), and (iii) end-to-end key-value pair extraction. However, existing evaluations often under-model two factors: heterogeneous and incompletely known key representations, and OCR-induced noise. This makes it difficult to assess model robustness in real-world settings. We present MedStruct-S, a benchmark specifically designed to evaluate these tasks under unknown keys and OCR noise. MedStruct-S contains 3,582 annotated real-world clinical report pages. Using MedStruct-S, we benchmark two representative paradigms: encoder-only sequence labeling with post-processing and decoder-only structured generation, covering four encoder-only and five decoder-only models spanning 0.11B to 103B parameters. Our results show that encoder-only models achieve the best performance for non-null-value key-conditioned QA despite being substantially smaller than decoder-only models. When comparing models of similar order of magnitude, encoder-only models still perform better overall. Without controlling for model scale, fine-tuned decoder-only models deliver the strongest overall results. These findings show that the benchmark provides a reliable and practical basis for selecting and comparing models across different semi-structured IE settings.

In classification tasks, models may rely on confounding variables to achieve strong in-distribution performance, capturing spurious features that fail under distribution shift. This shortcut behavior leads to substantial degradation in out-of-distribution settings. Task arithmetic offers a potential solution by removing unwanted signals via subtraction of secondary model updates, but it typically requires full fine-tuning, which is computationally expensive. Prompt tuning provides a parameter-efficient alternative by adapting models through a small set of trainable virtual tokens. Task arithmetic on the resulting prompts presents an appealing alternative to operations on entire models, but the extent to which this approach can limit reliance on spurious features remains to be established. In this work, we study whether composing soft prompts through task arithmetic improves robustness to confounding shifts. We propose Hybrid Prompt Arithmetic (HyPA), which combines task prompts with linearized confounder prompts to counteract spurious correlations. Across multiple benchmarks, HyPA consistently improves the robustness-performance trade-off relative to prompt-arithmetic baselines under distribution shift. We further analyze how HyPA affects hidden representations and find evidence consistent with it mitigating confounding either by reducing the influence of confounder signals on predictions or by suppressing them in the representation. These results establish HyPA as a parameter-efficient and promising approach for improving robustness under confounding shifts in the evaluated setting.

While sentiment analysis is the staple of financial NLP, capturing the nuances of 'why' behind that sentiment remains a challenge. There have been attempts to address this by analysing investor emotions alongside sentiment; however, this does not provide the additional granularity required to understand the target of the emotion/sentiment. We address this by augmenting the StockEmotions dataset with semantically structured opinion graphs, which provide granular semantic depth to the existing sentiment and emotion labels. Using a declarative LLM pipeline, we augment the StockEmotions dataset with opinion graphs for each sentence, derived from 10,000 comments collected from StockTwits. In addition, we study the effect of introducing opinion semantics on baseline classifiers using Graph Neural Networks (GNNs). Our analysis demonstrates that incorporating opinion semantics improves classification performance across different emotional spectrums

While pre-trained Transformer models achieve high accuracy on in-domain sentiment classification, they frequently experience severe performance degradation when transferring to out-of-domain data. We hypothesize that this generalization gap is driven by reliance on domain-specific spurious tokens. After demonstrating that post-hoc-token-level attribution drift fails to predict this gap, we propose Attribution-Guided Masking (AGM), a training time intervention that dynamically detects and penalizes highly attributed spurious tokens during fine-tuning. AGM's core component is a gradient based attribution masking loss ($\mathcal{L}_{mask}$), which can optionally be combined with a counterfactual contrastive loss to enforce domain-invariant representations, all without requiring target-domain labels or human annotation. Evaluated in a strict zero-shot transfer setting across four diverse domains with eight random seeds, AGM achieves competitive generalization compared to five strong baselines on the hardest transfer (Sentiment140): $\Delta$ = 0.244 versus DANN (0.264), DRO (0.248), Fish (0.247), and IRM (0.238), while uniquely providing token-level interpretability into which features drive the generalization gap. Our qualitative analysis confirms that AGM suppresses attribution on domain-specific tokens such as @mentions, hashtags, and slang, shifting reliance toward domain-invariant sentiment markers. Our ablation study further confirms that attribution-guided masking is the critical component: removing it or replacing it with random token selection consistently degrades performance on difficult transfers.

Niche-domain Indic ASR – digit strings, currency amounts, addresses, brand names, English/Indic codemix – is under-served by both open-source SOTA and commercial systems. On a synthesised entity-dense Telugu test set (held-out by synthesis system), vasista22/whisper-telugu-large-v2 (open SOTA) achieves Entity-Hit-Rate (EHR) 0.027 and Deepgram Nova-3 (commercial) 0.16. We close this gap with a self-contained TTS<->STT flywheel: an open-source Indic TTS pipeline synthesises ~22,000 entity-dense Indic-English code-mix utterances at <$50 marginal cost, and a LoRA fine-tune on top of vasista22 achieves EHR 0.473 on the held-out test (17x over open SOTA, 3x over commercial), with read-prose regression bounded to +6.6 pp WER on FLEURS-Te. Cross-language: beta-Hi 0.337 (7x vs vasista22) and beta-Ta 0.543 (22x vs vasista22, 22x vs Deepgram); on Hindi where Deepgram has substantial entity coverage, the flywheel underperforms commercial. All three beta models fall below pre-registered EHR targets (0.75 for Te, 0.65 for Hi/Ta); we report honestly. A native-human-recorded sanity check (n=20 Telugu) confirms transfer to real speech (beta-Te EHR 0.516 on native vs 0.473 on synth). An EDSA-isolation ablation (LoRA on FLEURS-Te alone) yields EHR 0.020 on the same held-out, attributing ~100% of the gain to the EDSA corpus. We additionally report a language-conditional finding: vanilla Whisper-large-v3 has Telugu-specific Script Collapse (SFR 0.46-0.71) that a per-language LoRA corrects (SFR 0.81-0.97), but the recipe is contraindicated on Hindi and Tamil where vanilla SFR >= 0.98. Code, holdouts, predictions, EDSA corpus, and entity dictionaries are released open-source.

We study how Large Language Models (LLMs) process negation mechanistically. First, we establish that even though open-weight models often provide wrong answers to questions involving negation, they do possess internal components that process negation correctly. Their poor accuracy is due to late-layer attention behavior that promotes simple shortcuts; ablating those attention modules greatly improves accuracy on negation-related questions. Second, we uncover how models process negation. We consider two hypotheses: models could use attention heads that attend to the phrase being negated and suppress related concepts, or they could directly construct a representation of the entire negative phrase (e.g., representing "not gas" as a vector that promotes liquids and solids). We apply a range of observational and causal interpretability techniques on Mistral-7B and Llama-3.1-8B to show that models implement both mechanisms, with the "constructive" mechanism being more prominent. Combined, our work deepens the understanding of LLMs' internals, highlighting construction-dominant computations and the coexistence of competing mechanisms within LLMs.

Millions of users turn to AI models for their information needs. It is conceivable that a large number of user queries contain assumptions that may be factually inaccurate. Prior work notes that large language models (LLMs) often fail to challenge such erroneous assumptions, and can reinforce users' misinformed opinions. However, given the recent advances, especially in model's reasoning capabilities, we revisit whether large reasoning models (LRMs) can reason about the underlying assumptions and respond to user queries appropriately. We construct queries with varying degrees of presuppositions spanning health, science, and general knowledge, and use it to evaluate several widely-deployed models When compared to non-reasoning models, we find that reasoning models achieve a slightly higher accuracy (2-11%), but they still fail to challenge a large fraction (26-42%) of false presuppositions. Further, reasoning models remain susceptible to how strongly the presupposition is expressed.