Papers

Automatic speech recognition (ASR) for African languages is constrained by orthographic inconsistency, annotation artifacts, missing audio, speaker and domain imbalance, and evaluation procedures that differ from deployment. We present an end-to-end engineering study adapting NVIDIA Nemotron 3.5 ASR Streaming 0.6B to Kikuyu, Dholuo, and Kalenjin. Starting from a Kenyan Swahili-adapted checkpoint, we retain its cache-aware FastConformer RNN-T, prompt conditioning, and streaming decoder during full-parameter fine-tuning. The study covers corpus auditing, Unicode normalization, split checks, duration filtering, low-rate continuation, validation-based checkpoint selection, true-streaming evaluation, artifact preservation, and isolated serving. On internal, adaptively consulted evaluation sets excluded from gradient updates at context [56,13], selected Kikuyu and Dholuo models achieve 42.97% and 33.98% WER, respectively. Dholuo records 9.59% CER and 8.13% no-space CER under its frozen historical label policy; Kikuyu records 7.79% no-space CER. Kalenjin remains a work in progress: v1-v reaches 68.74% WER on a 2,411-row clean-v3 diagnostic subset excluding long-pause annotations, digit-bearing references, and targets shorter than three tokens. Its checkpoint selection used a mixed-source validation manifest containing test-origin rows, so the score is not an independent generalization estimate. We also report negative findings involving non-speech labels, short-utterance over-generation, boundary-sensitive WER, and cloud job-lifecycle failures. We make no state-of-the-art claim because the internal sets, repeated consultation, and normalization differ from public benchmarks. This work provides an auditable account of adapting a multilingual streaming model into language-specific systems without discarding streaming constraints.

Large language models leak parametric knowledge of what followed a historical date into decision tasks indexed by that date – not necessarily a lookup of the realized outcome, but knowledge of the period all the same. Existence is settled; what users lack is a cheap way to audit a given model. We present HindsightBench, a black-box audit protocol that profiles parametric hindsight in any time-indexed LLM decision task at probe-level cost (no backtests, no logprobs, no corpus access). It chains a four-arm date-manipulation matrix (revealed/date-only/masked/transplanted), dual memory probes (date recovery; outcome recall), and six metrics – trigger strength, transplant effect, post-cutoff placebo, recoverability, behaviorally effective cutoff, and recall-accuracy dissociation – with explicit gates where identifiability is data-dependent. Applied to 15 models from seven vendors on a 258-node vintage-correct macro panel, it yields three patterns: (i) the date-trigger reflex is not a scale phenomenon – it tracks training recency, though what installs it is not identified here: absent across every 2024 open-weight row where it is measurable, including a 70B tier with cutoff-aligned recall propensity, present in every tested 2026-generation model, and switching on within one vendor lineage (Qwen3 -> Qwen3.6) in the same MoE family at ~3B active; (ii) effective cutoffs span 22 months across vendors and precede vendor-reported dates by up to eight months, invalidating calendar-window placebos; (iii) results are not invariant to serving – BF16 serving of an FP8-referenced model breaks the trigger estimate's stability while AWQ-INT4 preserves it, and a provider-locked reasoning regime makes one probe non-convergent – so the protocol pins quantization and thinking regime as part of its contract. We release the panel, preregistrations, audit rows, transcripts, and one-command regeneration.

Efficient teamwork typically combines global coordination with parallel execution, a principle not yet fully reflected in unified Vision-Language Model (VLM)-based document parsers. Existing unified parsers process an entire page jointly but generate its output through a single token-by-token autoregressive trajectory, creating a sequential bottleneck that grows with document length. Such full-page sequential generation overlooks a key property of document parsing: layout must be analyzed globally, whereas block content can be parsed in parallel. Based on this observation, we introduce HPD-Parsing, which replaces full-page autoregressive generation with a Hierarchical Parallel Decoding paradigm. A main layout branch organizes the overall document structure and dynamically assigns block-level content decoding to concurrent branches, while progressive multi-token prediction (P-MTP) further reduces the decoding steps within each branch. Experiments on public benchmarks show that HPD-Parsing achieves 4,752 tokens per second, delivering $2.62\times$ the throughput of the fastest existing document parsing model and $3.06\times$ that of the vanilla autoregressive baseline, while maintaining competitive parsing accuracy. These results establish hierarchical parallel decoding as an effective alternative to full-page autoregressive generation, opening a new direction for efficient unified document parsing.

This comprehensive study introduces an advanced Artificial Intelligence for Indian Legal Question Answering (AILQA) system tailored to the Indian legal context. AILQA leverages a variety of embedding and generative models, including recent Large Language Models (LLMs), to address the unique challenges posed by the intricate and diverse nature of Indian legal texts and to enhance the accuracy and reliability of responses to legal questions. We conducted rigorous evaluations using both lexical and semantic metrics, enriched by expert legal feedback, to ensure relevance and accuracy. Our findings underscore the effectiveness of the Retrieval-Augmented Generation (RAG) paradigm in improving answer quality, particularly in complex legal domains. Additionally, we assessed performance on standardized tests such as the All India Bar Examination (AIBE), thereby providing a robust benchmark for practical applications. Under the study's evaluation protocol, some AI-generated responses received higher ratings than the available reference answers, particularly when they contained accurate and relevant supporting details. This finding is specific to the evaluated dataset and rating criteria and should not be interpreted as evidence that the models generally outperform qualified legal professionals. We also discuss the challenges encountered, such as the need for precise context and the risks of model hallucination, and propose directions for future research to further refine AI capabilities in the legal field. This study aims to pave the way for enhanced legal decision-support systems, making them more accessible and effective for legal professionals and the public alike.

Chain-of-thought (CoT) reasoning is widely used to improve both the performance and interpretability of large language models (LLMs), yet the generated reasoning may not faithfully support the final answer. We study this problem from a causal perspective, where a faithful CoT process should follow the chain $Z\rightarrow X\rightarrow Y$, with $Z$, $X$, and $Y$ denoting the instruction, reasoning chain, and final answer, respectively. In this process, the instruction should affect the answer only through the reasoning chain. However, conventional autoregressive LLMs condition answer generation on both the instruction and the CoT, which still allows a direct instruction-to-answer shortcut. To address this issue, we propose CASE, a framework that combines training-time causal alignment and inference-time structural enforcement. During training, CASE builds counterfactual-CoT, biased-instruction, and empty-instruction datasets, and applies selective-loss fine-tuning to strengthen CoT-to-answer dependence while suppressing instruction shortcuts. During inference, CASE masks direct attention from instruction tokens to answer tokens, preventing the model from bypassing the generated CoT. We provide an information-theoretic analysis showing how these components promote faithful chains. Experiments on three models and four benchmarks show that CASE achieves a 37% average per-setting relative improvement in overall CoT faithfulness over the strongest baselines, exhibits stronger cross-dataset faithfulness transfer, and maintains competitive average accuracy. Code is available at https://github.com/oddwang/CASE.

This paper proposes AI Tour Meeting, a group travel planning framework powered by multiple Large Language Model (LLM)-based agents. The agents are instantiated with distinct personas and collaboratively seek an itinerary that satisfies their constraints and preferences through natural language discussion. The framework enables easy and flexible orchestration of such discussions by providing interfaces for configuring agent personas, discussion workflows, monitoring, and LLM deployment. Its primary use case is a simulation tool for analyzing the behavior of multiple LLM agents during tour planning discussions. This paper demonstrates the utility of the framework by presenting system validation and several analytical results obtained by the framework.

Researchers, managers and policymakers are exploring different approaches and data sources to map the development and the diffusion of Artificial Intelligence (AI). In this research note, we illustrate the opportunities offered by trademark data. We argue that AI trademarks can complement AI patents to capture different dimensions of AI innovation. AI trademarks can reveal the extent and ways in which companies exploit AI technologies to develop new goods and services. Importantly, trademark data offer a timely and globally available data source that covers all economic sectors. We present insights from using AI trademarks in an empirical exploration of Italian firms. In our discussion, we reflect on how AI trademarks can be used at different levels of analysis to tackle emerging questions about the development and diffusion of AI.

Large language models (LLMs) have driven rapid progress in electronic design automation (EDA), yet their application to radio-frequency (RF) circuit design remains limited by the scarcity of domain-specific datasets and standardized benchmarks. We present RF-Agent, which addresses this gap through textbook-driven knowledge distillation. A multi-agent Question-Thinking-Solution-Answer (QTSA) pipeline converts a subsection-level corpus from seven canonical RF textbooks into the first-of-its-kind RF-domain reasoning dataset (over 11,000 samples) with a dedicated multiple-choice benchmark. On this benchmark we study two adaptation strategies: supervised fine-tuning (SFT) and three retrieval-augmented generation (RAG) configurations (semantic, keyword, hybrid). Across multiple LLM families, domain-specific SFT significantly improves RF reasoning, especially for small and medium-sized models; among RAG configurations, semantic retrieval performs best, indicating embedding-based context alignment suits RF reasoning better than naive fusion. The dataset and benchmark provide a reusable foundation for future work on LLM-aided RF circuit design.

The deployment of Small Language Models (SLMs) in educational settings offers significant advantages in terms of privacy, cost, and scalability. However, SLMs often struggle with complex vision-based tasks, such as grading handwritten student exams, due to the high computational cost of processing large images and the visual distractions present on a full page. In this paper, we investigate whether cropping student responses using bounding boxes can improve the accuracy and computational efficiency of SLMs on a short-answer grading task. Using a dataset of scanned handwritten responses from the 2025 Australian Physics Olympiad, we evaluate the performance of several models ranging from 4B to 72B parameters under varying conditions of Chain of Thought (CoT) prompting and image cropping. Our results demonstrate that using bounding boxes significantly improves grading accuracy and reduces computational cost (FLOPs) across models. We conclude that bounding boxes are a crucial pre-processing step for deploying SLMs in large-scale, vision-based educational assessments.

Public institutions hold large volumes of sensitive documents and support tickets that cannot leave the premises, ruling out cloud-hosted language models entirely. We report on RAGAL, a retrieval-augmented assistant for the technical-support team of AFIR, the Romanian Agency for Financing Rural Investments, built and operated under three hard constraints: zero data egress (no external API calls, even for synthetic data), a read-only mandate (the assistant drafts, humans execute), and a single 8 GB consumer laptop as the only development and training machine. Over a Romanian-language corpus of ~25,000 chunks – 15,073 resolved support tickets and internal normative documents – we show that the highest-leverage investments were retrieval engineering and retriever fine-tuning rather than a larger generator: hybrid dense-sparse retrieval with intent routing raised our internal evaluation from 62% to 81%, and fine-tuning the bge-m3 embedder on real ticket data improved recall@10 from 0.663 to 0.850 (MRR 0.489 to 0.684) after 72 minutes of training. We document a general pitfall: single-domain fine-tuning silently degraded retrieval on the untouched document domain below the stock baseline, detected only after building a per-domain evaluation set and repaired with locally generated queries (GenQ). We report two counter-intuitive findings – PII masking improved generation quality, and a structural "anchor distillation" scheme made SQL hallucination impossible by construction – along with a reproducible recipe for full embedder fine-tuning in 8 GB of VRAM. Finally, since zero egress also rules out a cloud judge, we describe a substitute: a 744B-parameter model run on CPU, too slow to serve interactively but affordable in overnight batch, used as a second opinion whose limits we quantify. We release the sanitized pipeline scripts for institutions facing similar data-locality constraints.

LLM agent failures are difficult to debug because the step where an error surfaces is often not the one that caused it. Existing observability tools replay execution traces but provide little support for identifying the root cause or translating diagnosis into recovery. We present AgentDebugX, an open-source debugging framework that organizes debugging as a closed loop of Detect, Attribute, Recover, and Rerun. At its core, DeepDebug performs multi-turn root-cause diagnosis through global trajectory understanding, structure-guided investigation, and cross-examination. On the Who and When benchmark, DeepDebug achieves the best strict attribution accuracy among the evaluated methods on both tested open-weight backbones, reaching 28.8 percent exact agent-and-step accuracy on qwen3.5-9b versus 21.7 percent for the strongest single-pass baseline. On GAIA, DeepDebug repairs 13 of 73 failed tasks in a single rerun, compared with 4 to 6 for three decoupled self-correction baselines, improving overall accuracy from 55.8 percent to 63.6 percent. AgentDebugX exposes this workflow through a Python library, CLI, web console, and installable agentic skill, and provides an opt-in Error Hub for sharing scrubbed failure-diagnosis-repair bundles and reusing them as debugging memory.

Translating brain signals into text could restore communication for people with severe paralysis, yet practically usable systems to date rely on invasive electrocorticography (ECoG). Electroencephalography (EEG) offers a non-invasive alternative, and EEG-to-text (EEG2Text) has been widely explored. Interestingly, however, EEG2Text models generally rely on teacher-forcing evaluation; without it, they fail to generate meaningful decoding. This reliance prevents EEG2Text from being applied in real-world, non-academic settings. This has fueled numerous debates about whether EEG2Text is a meaningful direction, by extension, and whether EEG truly contains decodable linguistic information. Here, using a neuropsychology-informed paradigm, we find that existing EEG2Text benchmarks have neglected EEG instability, a flaw that has confounded inference and sparked debate. Our experiments furnish key evidence for the feasibility of teacher-forcing-free EEG2Text decoding. Accordingly, we assemble the Corpus OF Eeg-To-Text (COFETT) using a 128-channel high-density EEG cap, providing a benchmark dedicated to evaluating EEG2Text models. In comparisons with multiple existing benchmarks, COFETT achieves SOTA ability to distinguish among model performances and enables robust, teacher-forcing-free evaluation, thereby opening a path toward practical EEG2Text applications. COFETT is open sourced in https://github.com/baoyudu/COFETT.

Recent work extends Transformer residual pathways along two complementary axes: historical retrieval selects information from earlier depths, whereas multi-stream methods maintain multiple residual trajectories. These capabilities have largely been studied in isolation, and assigning an independent retriever to each stream still prevents one trajectory from influencing depth selection in another. We propose Dual Attention Residuals (DAR), which brings multi-stream interaction into historical retrieval through reciprocal cross-stream addressing. For each target stream, DAR computes depth weights from normalized states in the opposite stream and applies them to values from the target stream's own history. The retrieved states are combined for an unchanged Transformer branch and updated through constrained gated writes; a block-form variant operates on block-level histories to control overhead. Across dense models from 0.1B to 1B parameters and a 7B sparse-MoE model, DAR consistently improves validation loss over standard residual Transformers and Attention Residuals. Routing ablations show that the gain cannot be explained by an additional stream or value projection alone. Representation and intervention analyses further show that reciprocal cross-stream selection preserves depth-wise diversity and avoids the redundancy or functional imbalance observed in alternative two-stream designs.

Large Language Models (LLMs) are increasingly fine-tuned for critical-domain Question-Answering (QA), yet choosing which small model to adapt, before paying the cost of adaptation, remains difficult. Fine-tuning can improve domain alignment, but it may also erode prior knowledge, weaken instruction-following, or increase hallucination, especially when labeled data are scarce or rapidly evolving as in cybersecurity. We present FiT (Find before Fine-Tune), a task-oriented diagnostic framework that characterizes small LLMs along three capabilities required for cybersecurity QA: vocabulary recognition, parametric knowledge, and contextualization of retrieved information. Using FiT, we conduct an empirical study of five open-weight 7-billion-parameter models under two fine-tuning regimes. We find that fine-tuning does not uniformly help: it consistently degrades vocabulary and parametric knowledge in small models, and the two regimes trade off differently. Knowledge-focused tuning causes moderate, rank-preserving degradation, whereas instruction-focused tuning collapses measured knowledge through induced abstention, inverting the knowledge ranking while leaving retrieval-grounded contextualization essentially intact. We quantify these regime-specific patterns with rank-correlation analysis and show that pre-fine-tuning FiT scores anticipate the direction of post-tuning change. Our results suggest that task-oriented diagnosis can screen out unsuitable models, avoid unnecessary fine-tuning, and support safer deployment of small LLMs in cybersecurity QA pipelines.

Asynchronous reinforcement learning improves throughput by decoupling rollout generation from optimization, but the resulting staleness is an inevitable byproduct, compounded jointly by policy lag, engine delays, and mixture-of-experts routing. From a trust-region perspective, this mismatch is critical: in the finite-horizon improvement bound, training-inference divergence governs the approximation error, whereas PPO clipping only gates sampled outward updates and therefore acts as a sampled surrogate rather than a full-policy constraint. As a result, the high-staleness update can remain weakly controlled in exactly the asynchronous regime where stale rollouts matter most. We introduce the Staleness-Adaptive Trust Region (SAT), which uses the detached sampled log-ratio as a practical staleness proxy, identifies the high-mismatch tail within each batch through Staleness-based kernel function scaling, and contracts only the sign-selected endpoint of the nominal PPO interval using Effective contraction factors. This design preserves the baseline behavior on ordinary tokens, while making the update more conservative exactly on newly intercepted outward bands. We evaluate SAT in a fully decoupled asynchronous reinforcement learning setup built on Qwen3-30B-A3B-Base, leveraging SGLang as the inference engine and Megatron as the training pipeline. In this setting, SAT-GSPO w/ R3 attains the best observed AIME24 avg@8, reaching 35.83 at lag 1 and 34.79 at lag 8, while SAT-GSPO reaches 34.17 at lag 1. More broadly, the results indicate that aligning the clip interval with observed staleness heterogeneity is an effective way to stabilize the reported asynchronous regime.

Stance detection aims to identify whether a text expresses a favorable or opposing attitude toward a given target, and serves as an important task for various downstream applications. Although existing studies have achieved strong performance in monolingual settings, especially in English, many low-resource languages such as Catalan still lack sufficient annotated data for training effective models. Cross-lingual stance detection alleviates this problem by transferring stance knowledge from resource-rich languages to low-resource languages. However, most existing methods mainly rely on semantic alignment between texts and targets, while ignoring the reasoning process required for reliable stance inference. Although Large Language Models provide strong reasoning ability, their high computational cost and inference latency limit practical deployment. To address these limitations, we propose a rationale-guided knowledge distillation framework for cross-lingual stance detection. Specifically, we use Chain-of-Thought prompting to guide Large Language Models in generating informative rationales, and distill the resulting reasoning knowledge into a compact student model. We further design a dual-path distillation mechanism to align rationale-enhanced and rationale-free representations, together with their prediction distributions. In addition, two contrastive learning strategies are introduced to improve stance discrimination. Experiments on multilingual benchmarks demonstrate that our method consistently outperforms competitive baselines.

Progresses have been made on understanding emotion mechanisms of large language models (LLMs). However, how to explain emotion in LLMs, or even what constitutes good explanations, are less clear. Emotion representations, components, circuits are widely recoverable, but as explanations of a model's own computation they are circular; the emotion space dimensions tend to be arbitrary and non-terminating. A pressing question to ask is whether a more primitive set of internal variables does the work: the semantic primes of the Natural Semantic Metalanguage (NSM). Across four instruction-tuned LLMs (Llama-1B, Gemma-2B, Gemma-9B, OLMo-7B), experiments show that the NSM primes are (1) recoverable internal elements; and (2) on the reference model, intervening with a prime based direction controls emotion about three times as strongly, and twice as selectively, as the best appraisal based direction; and (3) the model treats a prime based explication as interchangeable with the corresponding emotion. These evidences suggest that NSM primes seem to be better explanans for emotion in LLMs than many alternative options according to scientific explanations criteria.

A single embedding space that covers text, images, video, and audio lets one index serve every query a user can pose. Embedding models built on vision-language backbones now lead text/image/video retrieval benchmarks but lack audio entirely, while audio-text retrieval is led by specialist systems that serve no other modality. We present the Fusion Embedding family, which adds audio to a frozen vision-language embedding base whose parameters are never updated: generation 1 (fusion-embedding-1) trains only a 16.4M-parameter connector between a frozen audio tower and the frozen base, and generation 2 (fusion-embedding-2) adds modality-gated deep adapters (44.2M parameters) whose branch never executes on text, image, or video inputs: their outputs are bit-for-bit those of the released base, verified after every training run. Because the base already binds text, images, and video, aligning audio to text alone makes audio-image retrieval emerge, with zero paired audio-visual training data. Alongside the recipe we map its design space with controlled negative results (rewriting training captions with an LLM, substituting a leaderboard-stronger audio tower, and widening the connector each reduce retrieval) and with training-protocol findings that we expect to transfer to any frozen decoder-LM embedding backbone. Both generations train in hours on a single GPU. Weights, code, and the evaluation harness are openly released.

Safety interventions on dual-use knowledge typically choose between destroying hazardous content (e.g., unlearning, filtering) and suppressing it at the output layer (e.g., refusal training); both pay a tax in adjacent-domain competence or over-refusal. We argue that the right operation is conditioning, not reduction: we show that hazardous knowledge can be retained in the model and behaviorally gated by a privileged control token. Our method, Token Inoculation, introduces a binding-and-branching approach. First, during continued pre-training, we mark hazardous content by inserting a special token alongside dual-use documents, so the model binds the marker to the underlying semantics of the hazardous domain. Second, during supervised fine-tuning, we teach the model to answer hazardous queries correctly when the special token is present and to refuse them when it is absent, thereby enabling selective refusal without removing dual-use knowledge. On hazardous domain (e.g., WMDP-Bio), Token Inoculation reduces accuracy from 79% to 18% while retaining 93% of the base-model's benign-domain performance (e.g., MMLU), achieving the best safety-utility trade-off against unlearning and refusal-tuning baselines across 1B-14B model scales. We further show that refusal selectivity is controllable through the quality of the conditioning signal and that domain-specific semantic binding during pre-training is critical for the conditional behavior to generalize beyond memorized triggers. Our results suggest that safety alignment is better cast as a conditioning problem than a forgetting one: behavioral control is more precise when sensitive knowledge is retained under controlled access than when it is destroyed.

Multi-vector dense retrieval models, such as ColBERT, achieve strong retrieval effectiveness by modelling fine-grained token-level interactions between queries and documents. Methods such as PLAID use centroid-based quantisation of each token's vector to reduce the index size and speed up retrieval while maintaining strong effectiveness. In this work, we introduce PLAID-PRF, a method that performs Pseudo-Relevance Feedback (PRF) over PLAID to reformulate ColBERT's query vectors based on the top-retrieved results. In contrast with prior methods that perform PRF on multi-vector retrieval models, PLAID-PRF keeps computational costs low by leveraging the internal PLAID centroid vectors, treating them similarly to tokens in traditional PRF methods. The method selects a small and diverse set of high-utility expansion vectors and appends them to the original query, rerunning PLAID to refine both candidate generation and final scoring. Extensive experiments on the standard in-domain MSMARCO and four out-of-domain BEIR benchmarks show that PLAID-PRF consistently improves retrieval effectiveness over various baselines. In particular, PLAID-PRF improves over PLAID by up to 4.3% nDCG@10 and 7.3% MRR@10, while introducing substantially less computation overhead than prior PRF methods. The results demonstrate that our proposed centroid-aware PRF method offers an effective and lightweight mechanism to improve the quality of top-ranked retrieved results. Overall, this work enables effective and efficient feedback-aware late-interaction retrieval without expensive query-time document-token clustering.

Latent-reasoning looped language models (LoopLMs) offer a different scaling path for machine translation (MT): instead of increasing parameter count or emitting explicit chain-of-thought tokens, they spend additional recurrent computation inside hidden states. We introduce LatentMT, the first systematic study of latent-reasoning LoopLMs for machine translation. LatentMT adapts a small 2.6B-parameter backbone model with lightweight training. Across 32 translation directions spanning high-, mid-, and low-resource languages, LatentMT achieves performance comparable to models three to five times larger. It is competitive in a high-resource language and achieves state-of-the-art performance on both mid-resource and low-resource languages. Studying the behavior of scaling the number of recurrent reasoning steps, we find that recurrent computation consistently improves translation quality in early steps, then saturates quickly afterwards. Our mechanistic analysis shows that hidden-representation differences shrink along the recurrent reasoning-step axis, supporting the observed saturation in performance. Finally, our efficiency analysis shows that LatentMT requires lower training and inference compute than much larger non-latent-reasoning models with similar performance, making latent recurrent computation a promising path toward compact, efficient, and strong machine translation.

Machine unlearning for vision-language models (VLMs) remains underexplored. Unlike language models, VLMs combine a language backbone with visual components, which makes unlearning more complex. There is a surprising phenomenon when moving from single-modality unlearning to VLM unlearning: a target forgotten by the standalone language backbone can still be recovered when image information is given to the full VLM. This shows that text-only feedback is not enough for reliable VLM unlearning. Motivated by this observation, we propose Stochastic Meta-Unlearning (SMU), a bilevel framework that uses VLM-level feedback to learn an unlearning-ready initialization. In the inner loop, SMU applies a few unlearning steps to the language backbone using text data. In the outer loop, SMU recomposes the updated backbone with the frozen VLM and evaluates forgetting and utility at the VLM level. This design makes the unlearning update aware of the final multimodal behavior, while still keeping the update local to the language backbone. Experiments on two VLMs, two multimodal meme datasets, and three baselines show that SMU achieves the best overall forget-retain trade-off. Compared with the strongest baseline for each metric, SMU reduces average Forget accuracy by 10.52 points and improves average Retain and Test accuracy by 20.10 and 17.01 points, respectively. More importantly, SMU also transfers to new forgetting targets and to different meta-test unlearning methods. These results suggest that VLM-level feedback can make language-backbone unlearning more reliable and more transferable for VLMs.

We present AutoIndex, a framework for learning representation programs: executable transformations that map raw documents into the representations exposed to a retrieval system. Rather than tuning retrievers, rerankers, or a small set of preprocessing hyperparameters, AutoIndex searches over programs that slice, enrich, normalize, reweight, or reorganize documents before indexing. At each iteration, AutoIndex performs validation-guided program search, in which agents diagnose failures of the current program and synthesize candidate updates, retaining only updates that improve retrieval quality under the resulting index. We evaluate AutoIndex on CRUMB, a benchmark of heterogeneous retrieval tasks, with BM25 held fixed across all experiments. The learned programs improve recall over a static full-document BM25 baseline on all 8 tasks, with average gains of +8.4% in Recall@100 and +8.3% in nDCG@10, and largest gains of +30.5% in Recall@100 and +43.6% in nDCG@10. These results suggest that document representation should not be treated as a fixed preprocessing choice made before retrieval begins, but as an explicit optimization target. Code to reproduce our results is available at https://github.com/auto-index/autoindex.

Misinformation detection tools often rely on binary true and false classifications or models trained on historical examples, limiting their usefulness when novel misleading narratives emerge. Here, we present FakeSpotter, a content- and strategy-agnostic tool designed to estimate the viral misinformation risk of textual content by measuring structural fingerprints of misinformation rather than directly adjudicating truthfulness. FakeSpotter operationalizes a theory-driven framework across linguistic, narrative, logical, and critical-thinking dimensions, using repeated LLM assessments and domain-specific logistic regression classifiers for short and long texts. In a labelled corpus of 764 texts from social media and FakeNewsNet, FakeSpotter achieved macro F1 scores of 0.788 for short texts and 0.793 for long texts on a held-out test set. FakeSpotter's interpretive layer provides explainable outputs through feature-based scores, signal agreement, and a caution index, and can be used for social listening. These findings suggest that identifying the structural fingerprints of misinformation can support early, explainable, and human-supervised assessment of potentially viral misinformation.

Generative AI (GenAI) applications have achieved rapid consumer adoption, yet little large-scale research examines user-perceived quality, trust, and adoption barriers. We present one of the first cross-application analyses of app store reviews for six major GenAI applications (ChatGPT, Gemini, Microsoft Copilot, Claude, DeepSeek, and Perplexity), comprising 17,012 English-language reviews from Google Play and the Apple App Store. We combine BERTopic topic modeling with RoBERTa sentiment classification and evaluate cross-application differences using chi-square, Kruskal-Wallis, and multinomial logistic regression with Bonferroni correction. Both components are validated against human coding using a stratified sample of 300 reviews. Results show that negative sentiment concentrates in advertising (91%), authentication (89%), server reliability (83%), and subscription pricing (73%). Sentiment differs significantly across applications, with Claude exhibiting the highest negative sentiment (47.7%) alongside a strongly enthusiastic user base, indicating statistically significant polarization. These findings are robust despite unequal review counts across applications. As exploratory observations, a subset of DeepSeek reviews raised geopolitical and data privacy concerns related to its Chinese origin, while a proposed Trust Friction Score summarizes application-specific trust and usability barriers into interpretable dimensions. The study provides validated and actionable evidence on user trust, usability, and adoption barriers in consumer generative AI applications.

Large language models (LLMs) encode rich stylistic structure in their hidden activations, but discovering which stylistic dimensions are salient for a given prompt typically requires supervised contrastive data. We present a training-free, prompt-conditional alternative: we repeatedly sample completions of a single prompt at elevated temperature, apply Principal Component Analysis (PCA) to the pooled hidden activations, and label the resulting axes automatically from the pole generations. We validate the discovered axes against 245 human-elicited stylistic annotations in a two-phase study. On our strongest model (Qwen-3.5-4B-Instruct), the top two axes match spontaneously requested human dimensions with 72.8% precision and 43.6% macro-recall, and 75.6% of validity ratings judge the axes' polar generations accurate to their labels, with 90.9% adjacent inter-annotator agreement. Discoverability is strongly model-dependent: both Qwen models and Llama-3.2-3B expose human-salient axes, while DeepSeek-7B-Chat drops to 35.3% precision, its leading components dominated by structural rather than stylistic variance. Simple PCA over a model's own decoding variance is thus an effective, low-cost probe of stylistic structure in LLM representations, one that also exposes sharp cross-model differences in how that structure is organized.

Subliminal learning shows that language models can transmit a hidden trait through outputs that appear unrelated to it. One proposed explanation, token entanglement, links animal and number tokens through the model's output vocabulary. Yet existing measurements answer different questions: whether outputs co-vary, fixed output vectors align, an answer can be read from a hidden state, or that state causally controls the answer. We measure each separately in a fixed animal-number prompting protocol. From Llama-3.1-8B to 70B, fixed output-vector similarity predicts behavior less well: the paired mean correlation change is -0.080 (95% CI [-0.127, -0.035]). A fixed output-head readout shows no resolved change in normalized depth AUC. To test control, we copy the temporary answer-position state from one number prompt into another at five depths and measure which prompt the final animal score follows. Donor-control AUC rises from 0.254 to 0.540, a paired change of +0.286 (95% CI [+0.272, +0.300]), with increases for all 18 concepts. The contrast remains with exactly eight transformer blocks remaining, while specificity and identity controls remain small or exact. In two Qwen models, scoring every digit in sequence does not recover the positive one-token association. Per-token averaging instead creates a positive pooled association that disappears after controlling number width, revealing a length confound. Thus, fixed geometry, observational readability, causal timing, and multi-token measurement are distinct properties of this frozen prompting channel. They constrain token-level explanations but do not identify the mechanism of training-time trait transfer.

Sampling can increase response diversity without producing history-dependent behavior. We formalize a different design target, structured unpredictability, as conditional dependence between an output and a persistent hidden state beyond what an observer can infer from the transcript. A selection layer updates a low-dimensional style-and-attention state from a capacity-limited stream, generates several responses with a fixed base model, and selects for novelty and state affinity. Evaluation uses scripted sequences of independent prompt turns: the base model receives the current turn and rendered state, but not the preceding dialogue; cross-turn dependence resides in the wrapper state and response selector. A synthetic implementation validates the pipeline and matches four prospectively hash-frozen divergence features at point level. In the final real-model grid (mlx-community/Qwen2.5-1.5B-Instruct-4bit; 56 sequences per arm), the mechanism increased lexical novelty over the low-variance and consistency-only controls by 0.073 and 0.023, respectively. Its stylometric-consistency contrast with novelty-matched sampling was equivalent to zero under the registered smallest-effect rule, so the joint novelty-consistency criterion failed. The original two-part accumulation criterion also failed; a revised final-grid contrast, frozen after the powered grid, found higher consistency than the memory-reset ablation (0.028, 95% CI [0.018,0.039]), but does not establish path dependence. Twin separation was not established (0.003, 95% CI [-0.011,0.019]); the mean curve's saturating curvature matched the frozen prediction, which without separation does not support path dependence. Probe-level capability equivalence held within +/-0.10 on a near-ceiling battery, while output quality was not evaluated. All outcomes are machine-scored; no claims about perceived mind or consciousness are tested.

Candidate-based decoding typically selects a completion for each prompt independently, but many applications require a collection of outputs that satisfies global, non-decomposable requirements. We formulate this setting as structured multi-prompt, multi-completion selection: given a candidate pool for every prompt, select one completion per prompt to optimize a collection-level objective. We instantiate the problem in character dialogue, where each character should remain consistent across situations, each line should fit its situation, and characters and situations should remain distinguishable. Our method, TalkMatrix, generates multiple candidates for every character–situation pair and jointly selects a complete matrix using four embedding-based consistency and diversity objectives. Because a weighted sum can improve some dimensions by sacrificing another, TalkMatrix maximizes the worst-performing objective through a two-level minimax formulation. We approximately optimize the resulting discrete objective with multi-start coordinate ascent, and compare it with local, partial-matrix, and generic combinatorial search baselines. We run experiments on $50$ synthetic role-playing scenarios and $25$ curated board game scenarios where multiple characters interact in predefined situations. An LLM-as-a-judge rates matrix-level selection higher than random and independent cell-level selection baselines. These results show the value of structured selection for globally controlled dialogue generation, while our empirical validation remains specific to role-playing scenarios.

Web content extraction is essential for reliable LLM data pipelines, yet existing methods often struggle to jointly satisfy accuracy, scalability, and adaptability. General-purpose extractors can be applied broadly, but they are often brittle on publisher-specific layouts and richer extraction targets such as metadata, images, and tables. Direct LLM-based extraction offers greater flexibility, but incurs substantial cost and latency at scale, while manually engineered publisher-specific parsers can achieve high accuracy but require substantial human effort to build and maintain. We introduce PACE, an agentic framework for learning publisher-specific extraction configurations from representative pages and user requirements. During training, PACE uses LLMs to analyze page structure and aggregate reusable extraction patterns. At inference time, the learned configurations instantiate a fixed deterministic extractor template, enabling scalable extraction without additional LLM calls. Experiments spanning article-body, metadata, and multimodal extraction show that PACE outperforms scalable non-manual baselines while approaching the quality of manually engineered publisher-specific parsers. PACE achieves stronger extraction of article text, metadata, images, and tables, demonstrating that agentic configuration learning can automate publisher-specific extraction for LLM-ready page representations beyond article text.