Papers

Food access is multidimensional, but population-level measurement still relies heavily on geography because perceived dimensions of access are difficult to measure at scale. Here, we use 25,125 Google Maps reviews from 49 grocery stores in Raleigh, North Carolina, to measure five dimensions of food access: availability, accessibility, affordability, accommodation, and acceptability. We identify review topics with unsupervised topic modeling and assign them to access dimensions using zero-shot classification, with 85.4% agreement against manual coding. The resulting store-level measures capture distinct aspects of food access and reveal barriers that geographic proximity alone does not capture. Comparisons between nearby stores in the same chain further show that identical store policies can be perceived very differently across locations, consistent with food access reflecting the fit between residents and their food environment. Perceived food access also follows systematic socioeconomic and demographic patterns that broadly parallel, but do not replicate, those observed for geographic access. These results show that online grocery reviews can provide a scalable complement to geographic measures of food access.

Prompt-policy editing offers a practical way to improve agents that synthesize executable workflows without updating the underlying model. However, persistent prompt editing has two coupled properties. First, edit locality does not imply effect locality: an edit confined to one policy segment can ripple through downstream execution, altering behavior beyond the edited segment. Second, edit effects are composition-sensitive: edits that work in isolation can interfere after composition, causing one or both to lose their benefit or become harmful. Persistent adaptation must therefore support two distinct decisions: identifying where the policy should change from execution feedback, and determining whether the resulting edit remains safe to persist after composition. To address these challenges, we introduce RIPPLE (Replay-Informed Persistent Policy Localization and Editing), which separates where an edit is made from whether it remains safe after composition. It diagnoses failed trajectories, maps each actionable failure to a predefined policy segment, and restricts the correction to that part of the policy. RIPPLE then evaluates candidates against the same iteration-start policy to compare their isolated gains, before replaying promising edits after previously accepted updates to expose downstream effects and interactions. Only edits that remain safe under composition are retained. We evaluate RIPPLE on Flow-HO, a synthetic held-out benchmark for executable workflow synthesis. RIPPLE improves validation success by up to 23.1% and yields positive gains on two additional frozen language-model backbones, while maintaining edit efficiency and low execution cost. Targeted interaction analysis further demonstrates both properties: a segment-local tool-use edit changes downstream resource resolution and validation, while an edit beneficial in isolation becomes harmful after composition.

Consonants contribute unequally to whether a word is understood. Given the limited time available for therapy, ranking consonants by contribution to intelligibility helps prioritize intervention targets in motor speech disorders. However, measuring this contribution relies on perceptual studies that are difficult to scale. This paper presents a scalable method that measures consonant contribution using acoustic masking. We silence one consonant at a time in an isolated word and test whether an automatic speech recognition (ASR) model still recognizes the word. We define a consonant's contribution score as the proportion of its masked instances for which the word becomes misrecognized, which we refer to as the mask-induced misrecognition rate (MMR). We relate MMR to two linguistic factors previously reported to correlate with consonant contribution, namely phoneme frequency and functional load. We apply this analysis across four languages, English, Spanish, German, and Czech, using three ASR architectures, MMS (encoder-only), Whisper (encoder-decoder), and Qwen3-ASR (LLM-based). Using partial Spearman correlations, we find that phoneme frequency correlates negatively with MMR while functional load correlates positively. In other words, more frequent consonants are less disruptive when masked, whereas consonants carrying more lexical contrast are more disruptive. Further cross-language analysis shows that consonant rankings agree only partially across languages, indicating that consonant contribution is language-dependent.

Factual hallucination in closed-book question answering is often treated as a coverage problem: a model fails because the relevant fact is absent from its internal memory. This view misses a second source of error. Even when a fact has been observed, finite memory may force it to be stored only approximately. We study this effect through a simple coverage–compression model of factual recall. We consider an unstructured question-answering task with $N$ possible queries and $K$ possible answers. A learner observes $M$ training facts, compresses them into at most $B$ bits, and answers uniformly drawn test queries without retrieval. For a uniformly random ground-truth mapping, we prove $\mathcal{E} \geq \frac{M}{N}\delta^\star\!\left(\frac{B}{M}\right) + \left(1-\frac{M}{N}\right)\left(1-\frac{1}{K}\right)$, where $\delta^\star(r)$ is the inverse rate-distortion function of a uniform $K$-ary source under zero-one loss. The two terms separate compression distortion on observed facts from missing coverage on unobserved facts. The bound gives a compact way to reason about selective memory, forced compression, structure, retrieval, abstention, and long-context organization. We study the predicted signatures with theory-implied simulations and controlled fact-injection probes in modern language models that vary fact load and effective trainable memory. The result is not a complete theory of hallucination, but an information-theoretic account of a separable failure mode: lossy recall of observed facts under finite memory.

Development and humanitarian organizations produce and support surveys, administrative registries, and other data resources to inform research, policy, and operations, yet systematically identifying where these datasets are referenced remains difficult. Such references are dispersed across research papers, project documents, humanitarian reports, and other unstructured text, limiting both the ability to trace data use and to identify potential gaps in data availability or dissemination. We present a weakly supervised framework for adapting dataset extraction to forced displacement and Fragile, Conflict, and Violence (FCV) documents without first constructing a large manually labeled training corpus. A lightweight model trained on general research literature generates candidate dataset mentions from unlabeled domain documents, which a frontier large language model (LLM) reviews in context, validating or rejecting candidates and correcting their extraction boundaries. The resulting annotations are supplemented with targeted synthetic and contrastive examples and used to fine-tune the lightweight model for large-scale extraction. We evaluate the resulting model on an independent gold-standard benchmark of 1,706 text passages spanning research, humanitarian, and operational documents. Across the full benchmark, the model achieves 74.1% precision and 70.5% recall at the mention level; among passages containing dataset references, precision reaches 89.5%. At the passage level, the model achieves 88.2% accuracy and 88.6% specificity in distinguishing passages with dataset references from those without them. These results demonstrate a practical approach for constructing domain-specific supervision when labeled data are limited, and provide a technical foundation for larger-scale analysis of data use and potential gaps in the displacement data landscape.

Assistants built on large language models are expected to write in their users' own voice. Most systems summarise the user's preferences and include the summary in the prompt. This is the wrong way round. Preferences are only the surface of a person and change with the task, while the underlying personality stays the same, so storing preferences alone means relearning the user afresh whenever the task changes. This paper makes four contributions. First, we describe an effect we call personality seepage: the wording of a prompt carries traces of the writer's personality, which the assistant copies without knowing the writer. Second, we propose the Atomic User Model (AUM), a readable profile with a stable identity core surrounded by four layers covering psychological, cognitive, experiential, behavioral, and social details, plus notes on inner conflict and authenticity. Third, instead of inserting the entire profile, we use AUM as a searchable index, in which a task classifier, a selection step, and a budgeted retriever pass along only a few relevant fields. Fourth, we test the pipeline with 16 simulated users, 6 style-sensitive tasks, and 3 seeds. Eight retrieved fields matched the writing quality of the whole profile, while using only 23 percent of the context (211 tokens instead of 915). They scored 0.24 points higher than a plain preference note on a five-point scale. Accuracy in picking a user's own writing from four samples rose from 14.9 to 42.7 percent, where guessing gives 25 percent. Four pre-registered controls showed no effect, so the gain comes from the profile's structure rather than the search method. Personalization helps most for the users for whom a generic assistant imitates them the worst.

Checking Quran recitation from an ASR transcript requires distinguishing unresolved mistakes from repetitions, repairs, opening formulas and accepted spelling differences. We report a completed human annotation of 100 production recording cases: 348 scored units and 162 localized events across ten combined labels. An executable evaluator scores labels and word positions together. A plain diff reaches label-aware F1 0.525 and localization F1 0.826; adapted production cleaner/alignment components reach 0.518 and 0.786, with exact-span F1 0.505 for both. Correcting the adapter's word coordinates recovers all five annotated repetition events, showing why annotation interfaces must be checked before interpreting baseline failures. In a preliminary pilot, eight single 20-minute runs across three coding agents and eight models span label-aware F1 0.143 to 0.892: seven land far above every baseline, and one collapses below the naive diff from a missing normalization step. Across the six, 970 of 972 gold-event instances draw an overlapping prediction, so what remains is not detection but convention: span extent, and the labels whose boundary is stipulated by adjudication rather than visible in the text. Seven of 162 events defeat all six same-day runs, five of them one orthographic rule, and the strongest run still misses the same ones. No run annotated before building, so the pilot measures the algorithm half of the task only.

In this study, we examine whether a general shell can outperform specialized tools on enterprise tasks. Shell-based agents have shown strong results in coding, but enterprise work also involves moving between applications and services, coordinating with coworkers, and performing professional analysis. We compare five tool interfaces on TheAgentCompany and APEX-Agents using Opus-4.8 and GPT-5.5: typed tools, typed tools plus bash, bash alone, bash with persistent agent-synthesized tools, and programmatic tool calling (PTC), which runs programs whose actions are restricted to a typed tool catalog. Bash alone outperforms typed tools on both benchmarks, improving score by 21.8-24.5 pp on TheAgentCompany and 4.8-7.4 pp on APEX-Agents while using 19-72% fewer total tokens. Adding typed tools or persistent tool synthesis to bash produces no detectable pooled score gain. PTC uses fewer tokens than direct typed calls with broadly similar task performance, but generally underperforms bash alone in both quality and cost efficiency. For enterprise practitioners, these results favor bash alone when arbitrary execution can be isolated and PTC when security or compliance policies require a fixed tool catalog.

As the supply of human-written text is exhausted, it has become standard practice to repeat language model training data. Prior work has studied data repetition for densely activated Transformers, but the effects of data repetition remains largely unexplored for recently dominant sparse architectures such as Mixture-of-Experts (MoE), despite their increased compute efficiency. We vary data repetition rates across single- and multi-domain data mixes, and across MoE settings, including expert count and granularity. We consistently find, for models ranging from 80M to 1B active (8.5B total) parameters, that MoEs degrade more rapidly under data repetition. This effect increases with sparsity, dictated by total rather than active parameters. While 80M dense models can repeat data over 8x with minimal degradation, MoEs instead begin to suffer at 4x, and deteriorate rapidly, ceding their performance benefits in all-unique data settings to underperform dense models after 32x. We experiment with existing regularization methods as a potential remedy. We find that some methods, such as dropout, can mitigate overfitting. In particular, with strong masking-based regularization, MoEs are able to outperform dense models even when data is repeated more than 64 times. However, no method fully matches the performance of all-unique training data. Finally, we analyze internal mechanisms correlated with MoE overfitting in high repetition regimes, and find that MoE routing universally stabilizes early in training, and that expert specialization correlates with overfitting to repeated data. In sum, our work addresses the adverse interactions between sparsity and data repetition: we present evidence for the core mechanisms of overfitting and its potential remediation, and suggest promising avenues for future methods to reduce over-specialization in model parameters by disrupting memorization patterns.

Out-of-distribution length generalization, namely to extrapolate a task from short to longer context, has been studied intensively for transformers. Here we focus on distance generalization, which probes performance when inter-token distances are changed between training and inference, while keeping a fixed context length. We construct two synthetic delay copy tasks, both involving finite distances between source and recall, where tokens are copied either fully or selectively, and test models on delays unseen during training. We address three questions: (A) Do positional encoding schemes such as RoPE and ALiBi improve distance resolution relative to no positional encoding (NoPE)? (B) How does data diversity, the number of inter-token distances seen in training, affect performance? (C) When is distance transfer learning positive or negative? We present a thorough investigation, finding that it is paramount to improve our understanding of the underlying mechanisms.

Spatial reasoning depends not only on metric properties such as distance, angle, and shape, but also on topological relations that remain invariant under continuous deformation. Cognitive science identifies these relations as foundational to spatial understanding, yet foundation-model evaluations largely focus on metric or viewpoint-dependent relations. We introduce MindTopo, a benchmark of topological intuition across five properties grounded in cognitive science and formal topology: continuity, separation, order, enclosure, and knots. MindTopo evaluates each property at two cognitive levels. Reasoning asks a model to identify topological relations or infer how they change. Planning instantiates a foundation model as a closed-loop agent whose policy selects environment actions. MindTopo contains 11,030 instances across 13 procedurally generated task types with controllable difficulty. We benchmark 14 MLLMs and study agent configurations augmented with image and video generation, including 3 video generative models in planning settings. Every MLLM performs better on reasoning than on planning, and the best-performing model remains far below observed human performance. On Qwen3-VL-2B-Instruct, supervised fine-tuning and reinforcement learning improve reasoning more than planning. Generated observations retain local cues and reach plausible endpoints, but audited rollouts do not reliably follow environment dynamics or preserve topology across transitions. Our website is at https://mind-topo.github.io/

As Speech Large Language Models (speech-LLMs) become increasingly multilingual, Arabic remains significantly underrepresented, highlighting the need for dedicated infrastructure to train and evaluate Arabic speech-LLMs. To address this gap, we introduce Nuha-Speech, a comprehensive initiative to develop general-purpose Arabic speech-LLMs spanning dataset construction, model training, and systematic evaluation. Specifically, we constructed a large-scale Arabic Speech Question-Answering (SQA) corpus comprising over 1.5 million training samples to allow instruction tuning over a broad range of core speech tasks. Then, the corpus was used for supervised fine-tuning based on Qwen-Omni model variants at different scales. Finally, we designed an evaluation framework featuring diverse tasks and tailored metrics. Through this work, we aim to establish foundational infrastructures for Arabic Speech-LLMs under constraints imposed by limited Arabic speech resources.

Large language models generate fluent text that can contain unfaithful claims – a phenomenon known as hallucination. We present a multi-signal detection pipeline combining fine-tuned DeBERTa-v3 classification, Monte Carlo (MC) Dropout uncertainty quantification, and temperature-scaled calibration for response-level hallucination detection. Evaluated on the HaluEval benchmark, our pipeline achieves F1=0.915 and AUROC=0.977 on general-domain tasks, with per-task F1 scores of 0.97 (QA), 0.96 (Summarization), and 0.82 (Dialogue). MC Dropout inference further improves accuracy to 93.2%. A context ablation study confirms the model performs genuine entailment reasoning rather than exploiting surface patterns, with summarization F1 dropping 24% when knowledge context is removed. Learning curve analysis reveals that 25% of training data captures 77% of full-data performance. Beyond detection, we apply Direct Preference Optimization (DPO) to a Qwen2.5-0.5B generator, reducing its hallucination rate from 85.5% to 37.7% (55.9% relative reduction) as measured by our detector. Cross-domain evaluation on the SciFact biomedical benchmark shows that general-domain training transfers poorly (F1=0.52), motivating domain-specific fine-tuning. PubMedBERT fine-tuned on SciFact achieves F1=0.63 and AUROC=0.81, demonstrating that domain-matched pre-training is the strongest adaptation strategy. Code and models are available at https://github.com/varunteja99/hallucination-detection-nlp

Many biological discovery problems require experiments to be selected sequentially under constrained budgets. CRISPR screening is a prominent example, as exhaustive perturbation testing is often infeasible and candidate perturbations must instead be prioritized over multiple experimental rounds. Despite the importance of this problem, existing benchmarks for adaptive hit discovery remain limited in scale and diversity. Here, we introduce AssayBench-Loop, a large-scale benchmark for adaptive hit discovery comprising 1,389 CRISPR screens across five phenotype categories. Beyond enabling systematic evaluation, its scale makes it possible to learn acquisition strategies across historical experiments. Building on this resource, we introduce AssayLoop, a sequential experimental design framework combining AssayFormer, a transformer-based amortized acquisition policy trained across historical screens to adapt from experimental feedback, with LLM-derived biological priors through an adaptive handoff. In this view, completed experiments become training data for learning how accumulated evidence should guide what to test next, while LLMs provide prior biological knowledge to seed the search. We further introduce AssayLLM, showing that the same principle can be extended directly to an LLM through task-specific post-training. On temporally held-out screens, AssayLoop achieves a 5.67-fold enrichment over random selection and recovers 27.7% of hits after assaying approximately 5% of the candidate library, outperforming existing adaptive-design methods and standalone LLMs, and AssayFormer alone. Performance improves with increasing historical training data and transfers to phenotype categories excluded from training. These results demonstrate the value of learning acquisition policies across historical experiments and combining them with broad biological priors for efficient adaptive hit discovery.

Recursive self-improvement (RSI) enables AI systems to turn experience and feedback into persistent changes that improve both their capabilities and the process of future improvement. We first use the Headroom-Closed Index (HCI) to reveal the problems of existing LLMs, then introduce the RSI concept and its development roadmap: from improvement-execution autonomy, improvement-strategy autonomy, experience-acquisition autonomy, and environment-adaptation autonomy, to recursive meta-improvement. Next we examine RSI across scenarios (e.g., scientific discovery, embodied intelligence, software engineering), highlighting their distinct requirements and development speeds. Drawing on diverse industry practices and preliminary empirical evidence, we connect RSI research with practical systems and identify key challenges to achieving genuine RSI.

A language model normally begins training with random word embeddings: whatever 'banana' means must be learned from training corpora. I implement St. Augustine's picture of word learning, meaning by ostension, for a small masked language model (DeBERTa) trained on 10M words: before training, visually grounded tokens receive embeddings derived from the image regions they label; other tokens start random. Visual initialization leaves a measurable imprint that lasts until the end of training. At the same time, the effect remains invisible under most BabyLM benchmarks, which probe abstract grammatical knowledge: visual initialization does not affect performance there. The only zero-shot exception is object-property knowledge (COMPS), where seeding helps in every configuration. To follow up on this result, I build a corpus-tailored version of the Visual-Property Swap benchmark, which tests color, material, size, and shape knowledge, with per-item training frequency and seeded status. Here, vision-seeded models have a persistent, seed-replicated advantage. Function words and abstract vocabulary also receive strong visual seeds and retain them throughout training, and the training objective draws on them: held-out mask-prediction loss falls for these words in every seed. However, no benchmark I run registers this. What evaluation would pick this up remains an open question.

Truth and evidence-based communication provide important foundations for democratic governance, accountability, and collective decision-making. Prior work shows that evidence-oriented language in US congressional floor speeches has declined since the mid-1970s, alongside broader changes in legislative productivity and polarization. This study shifts the analysis from congressional sessions to individual members of Congress to examine whether epistemic orientation varies systematically across legislators and whether it relates to political behavior and legislative effectiveness. Using the Evidence-Minus-Intuition (EMI) score, we measure the relative prevalence of evidence-oriented versus intuition-oriented language in congressional floor speeches and Twitter posts. We link these measures to legislator-level data on ideology, institutional position, communication context, and Legislative Effectiveness Score (LES). The results show that more ideologically extreme members use less evidence-oriented language on the congressional floor. EMI also exhibits cross-platform consistency with members who use more evidence-oriented language in floor speeches also being more evidence-oriented on Twitter, although EMI is lower on Twitter overall. Finally, EMI in congressional speeches is positively associated with individual legislative effectiveness, even after accounting for ideology and extensive political, institutional, demographic, topical, and communication volume controls. These findings suggest that evidence-oriented language is not only an aggregate feature of congressional discourse but also a meaningful attribute of individual-level legislative communication and effectiveness.

Speech large language models (SpeechLLMs) offer reduced latency and retain paralinguistic nuances that are typically lost in cascaded automatic speech recognition (ASR) and text-based LM architectures. However, they continue to lag behind text-only LLMs on complex reasoning tasks, while real-time spoken interaction imposes strict latency constraints. Although prior works employ Chain-of-Thought (CoT) and concurrent reasoning to enhance reasoning capabilities without inducing prohibitive delays, an inherent accuracy-latency trade-off persists. In this paper, we investigate whether a streaming SpeechLLM can dynamically revise its reasoning traces on the fly. We introduce RetroThinker, a multi-stage post-training framework that equips the Moshi model to self-verify and forward-correct CoT steps during inference. RetroThinker combines supervised fine-tuning (SFT) on curated retrospective thinking data with length-based direct preference optimization (DPO) to optimize retrospective during early reasoning (i.e., reasoning concurrently while the user speaks). Evaluated on the GSM8K benchmark, RetroThinker significantly improves the accuracy-latency trade-off over non-retrospective baselines, achieving an 11% absolute accuracy gain at a comparable latency.

Language identification in code-mixed text, largely observed in social media, is highly essential when users frequently switch between multiple languages within a single utterance. Accurately identifying the languages of code-mixed tokens becomes an urgent necessity. Traditional language identification models, designed for monolingual text, are not well suited for token-level language identification in code-mixed settings. We formulate the task as a sequence labeling problem and fine-tune contextual transformer-based models MuRIL and XLM-RoBERTa best suited for Indian languages. We evaluate these systems on three different data configurations (Hindi, Gujarati, and Bengali) to predict language labels for individual tokens. We release a benchmark for language identification in code-mixed tokens with manually annotated test sets. We propose two approaches of code-mixed generation using parallel sentences of three languages. The trained models demonstrate the effectiveness of contextual embeddings for token-level language identification in multilingual social media text. For reproducibility and to facilitate future research, we publicly release our fine-tuned models.

Cardiovascular screening models trained on national health surveys routinely report areas under the receiver operating characteristic curve (AUROC) near 0.89. We asked whether that accuracy reflects learning or target leakage, whether tabular foundation models change the answer, and whether the properties deployment requires survive joint examination. We benchmarked ten classifiers spanning linear, tree-ensemble, neural, glass-box, and tabular foundation classes for prevalent myocardial infarction in 442,067 respondents of the 2022 Behavioral Risk Factor Surveillance System across five feature tiers of decreasing leakage risk. Each was audited for discrimination, calibration, fairness at an explicit screening threshold, conformal coverage, explanation faithfulness, and inference cost, then applied – models and thresholds frozen – to 430,755 respondents of 2023. Removing two post-diagnostic features cost every model 0.049-0.051 AUROC, collapsing the field into a 0.0045-wide band. The glass-box explainable boosting machine was non-inferior to every alternative within a pre-specified 0.005 margin while scoring the cohort roughly 104 times faster than the strongest foundation model. One threshold detected 75.4% of women's infarctions against 89.0% of men's; editing the model's shape functions reduced the gap to 0.010. Marginal conformal prediction gave 0.86 coverage to men and 0.82 to adults over 60; Mondrian calibration repaired every stratum. Frozen models transported within 0.002 AUROC. Reported headroom in this literature is a property of the feature set, not the learner. Transparency cost nothing measurable and made fairness repair and uncertainty conditioning directly auditable. Evaluation practice, not model capacity, is the binding constraint.

Large language models are often fine-tuned, shared, or downloaded from third parties, so a deployed model may carry a hidden backdoor that behaves normally on benign inputs but switches to attacker-controlled behavior when a secret trigger appears. While backdoors can be audited before deployment, runtime monitoring remains important for models that are frequently updated. The challenge is that LLM serving is latency-sensitive: existing inference-time detectors either rely on assumptions about the trigger form, which can fail on stealthy attacks, or require extra model computation, such as input perturbations or an additional generation pass. We introduce SpecGuard, an inference-time backdoor detector that repurposes speculative decoding at zero added model-computation cost. Speculative decoding speeds up inference by using a small draft model to propose tokens and a target model to verify them. We observe that this verification process already exposes a useful signal: when a backdoor is triggered, the target model shifts toward the attacker's behavior, while a clean draft model does not predict this shift, causing the draft-token acceptance rate to change. We formalize when this signal appears and show that an attacker who suppresses it must also weaken the backdoor. Across diverse backdoor types and model families, SpecGuard reliably detects triggered behavior, including stealthy cases where input-level filters are blind, while avoiding the extra generation cost of existing runtime detectors. Speculative decoding therefore doubles as a free, always-on signal for detecting backdoored LLM behavior.

Automatic speech recognition (ASR) systems and audio language models (audio LMs) now report low error rates on monolingual benchmarks, but their behavior on code switched speech in low resource, diacritic rich languages remains poorly characterized. We present a switch aware evaluation of eleven modern systems (six ASR models and five audio LMs) on English Yoruba code-switched speech, using a deterministic 2000 utterance evaluation set and a shared scoring pipeline. Beyond word error rate (WER), we report switch localized diagnostics: a switch entry token error rate (SETER), windowed switch point error rates, language specific error rates, and a diacritic insensitive WER. Our central finding is that aggregate WER hides code switching behavior. The best system by WER (an ASR model) is statistically indistinguishable from a leading audio LM on WER, yet the audio LM is significantly better on every switch localized metric. Across faithful systems, Yoruba token recognition collapses (error 0.97 for almost all systems) while English tokens are recognized far better, and errors concentrate sharply at switches into Yoruba. Several generative audio LMs fail as exact transcribers, producing translation, verbosity, and prompt leakage that are strongly prompt dependent. We release manifests, metric implementations, and evaluation scripts to support reproducible, switch aware benchmarking for African code switched speech.

Cross-cultural understanding has become increasingly important in today's highly connected, cross-national world. The success of LLM-based technologies is now driving the development of automated tools to aid understanding for nonnative people trying to succeed in cross-cultural environments. Building such automated tools is often done by leveraging in-thewild text, audio, and video data. This paper presents techniques for improving speech recognition-based transcript creation in multiple languages from videos to better train these automated tools. The focus is on processes and speech tools that can easily be used by cross-cultural tool builders without requiring deep speech processing expertise. Using publicly available videos from YouTube and Whisper-based tools, average transcription error rate across seven languages (Spanish, Japanese, Korean, Mandarin, Turkish, Russian, and Hebrew) of 30% are observed. With a modest amount of fine-tuning data, the average error rate can be reduced to 20% making such output much more usable for downstream processing. Speech and metadata associated with these videos that can be used by the community to further refine these experiments are released as well.

Large language models are superseded every few quarters; clinical evidence takes years. We asked whether medical research is keeping pace with the systems it evaluates. PubMed returned 11,628 records for January 2023 to June 2026 across fourteen clinical domains, growing 45-fold; 2.5% used a randomised, controlled or prospective design. Evaluation lag, from a study's newest named model release to its own publication, widened from 1.33 to 6.08 quarters. Because discontinued models age mechanically, we benchmarked this against a counterfactual holding model composition fixed: migration to newer systems offset only 56% of the drift (95% CI 50-65). Randomised trials evaluated models a median 4.6 quarters older than other designs (P = 3 x 10^-19), yet among studies naming a model still under development no design differed from any other; 62% of randomised trials evaluated a discontinued family. Rigour and currency are in tension, and that tension reflects model selection rather than research timelines.

Large language models (LLMs) are increasingly used to analyze and rewrite news, yet current framing studies mainly evaluate generation, detection, or whether rewritten text appears more neutral. They do not directly show whether a model can undo a known framing transformation while keeping the facts fixed. We introduce a controlled inversion test over three established textual realizations of framing: evaluative lexis, agency realization, and information salience. Across 60 news articles and three intervention strengths, this yields 540 paired variants with preserved atomic facts and recorded edits. Across Qwen, DeepSeek, and Kimi, factual preservation remains near 0.84, whereas intervention reversal is 0.044–0.068. Even when both framing type and direction are recognized correctly, pooled reversal reaches 0.071. These results reveal a clear separation between factual fidelity, framing recognition, and framing inversion: recognizing how an article is framed does not imply that the framing can be undone.

Per-token gating of forward/reverse KL losses has become a standard technique for on-policy knowledge distillation (OPD), but existing methods such as EOPD (Jin et al., 2026) and ToDi (Jung et al., 2025) each fix a single gating signal and a single gating direction, and the two have never been compared directly. We introduce a four-coefficient parameterization lambda_t = sigma(a * h_t + b * u(x) + c + d * gap_t) in which direction-aligned proxies of EOPD and ToDi appear as one-dimensional (1D) restrictions, and which adds multi-channel composition and an explicit bias as further degrees of freedom. On TweetEval (Barbieri et al., 2020) emotion and hate, with a Qwen3-32B teacher and a Qwen3-4B student, configurations in the full family reach higher accuracy than the matched-magnitude single-channel (entropy-only / gap-only) 1D restrictions in 33 of 36 comparable cells, and a 26-cell mean-match isolation experiment places dynamic gating ahead of effective-KL-matched static baselines in 19 of 26 cells. Because cells share training data, models, and parameter substructure, we report both counts as exploratory aggregate directional evidence rather than as independent hypothesis tests. Targeted three-seed paired replications of the nine headline comparisons singled out by that sweep – including a third task, offensive – are directionally consistent, but individually smaller than the single-seed estimates and not significant at n=3. We therefore present the parameterization primarily as a shared coordinate system for comparing per-token gating designs in short-output classification OPD.

Per-layer differential privacy (DP) clipping improves gradient fidelity in federated learning by allocating per-matrix clipping budgets proportional to parameter count. We show that this recipe breaks for speech large language models (speech-LLMs), when the acoustic encoder and the language decoder differ by an order of magnitude in update norm. Single-pool per-layer methods suffer cross-component budget collapse, dragging word error rate (WER) far from flat global clipping or collapsing training entirely. When the norm imbalance is milder, adaptive single-pool methods partially recover, confirming that collapse severity scales with the inter-component norm ratio. We empirically diagnose the root cause across six per-layer methods and three speech-LLM architectures. We then propose \emph{$\alpha$-split}, a two-pool allocation that normalises encoder and LLM parameters into independent pools, and show that joint $\ell_2$ sensitivity and the original $(\varepsilon,\delta)$-DP guarantee are unchanged. At architecture-calibrated $\alpha$, our method recovers WER utility compared to flat DP, while granting the encoder $4.47{\times}$ tighter per-component noise protection against speaker voice-based gradient-inversion attacks at only $+2.6\%$ LLM noise overhead.

Allowing large language models (LLMs) to retrieve information from a set of trusted documents can increase reliability and reduce hallucination. However, recent work has demonstrated that retrieval-augmented generation (RAG) can have unintended side effects on the overall safety of the generated responses, when prompted for harmful or dangerous content. A clearer understanding of the mechanisms leading to this result is needed, as increasing numbers of end users turn to RAG to incorporate corporate documents and knowledge bases into LLM-based systems. We introduce RAG-Safety-Bench, a benchmark to measure the safety impact of RAG on LLM models. By removing the confounding effect of retriever quality, and cleanly separating the problem into four conditions – non-RAG, RAG with an oracle document containing the answer to the harmful request, RAG with documents related to the harmful request but without the specific answer, and RAG with random, safe documents – the benchmark isolates the impacts of different factors in the observed safety degradation. We report results across five open-source LLMs, showing an inverse relationship between benign and unsafe capability, strong evidence that baseline safety guardrails do not lead to downstream safety guarantees in the RAG case, and model-specific support for previous findings that even benign documents can lead to unsafe generation in retrieval-enabled systems.

For industrial content risk control, the real deployment constraint is not average accuracy but how much risk can be auto-handled under high precision and second-level latency. We present SIRF (Spec-Internalized Risk Foundation Model), which internalizes a platform's complex policies, synthesized without additional human annotation via EntiGraph, MAGA rewriting and account-level chain-of-thought (CoT), into the weights via continued pretraining (CPT), so rules are applied at high precision under an ultra-low-latency, verdict-only deployment. A controlled same-source comparison (Qwen3-8B-SFT vs. SIRF-8B-SFT, identical policy injection and verdict-only output form, differing only in policy-grounded CPT) attributes the gain to internalization: SIRF-8B-SFT reaches 71.3% Black Recall@P95, +15.1pp over the baseline, using only ~70M CPT tokens without harming general ability, and among included, logprob-available models under this interface it matches or exceeds far larger systems. SIRF is deployed as a tree-model adjudication layer (20% more mis-penalized samples recovered) and transfers to a freezing scenario at low cost (~70% relative mis-penalization reduction).

Large language model serving costs scale directly with output sequence length, yet standard preference alignment often inflates response verbosity without improving utility. We study whether the parameterization of post-training updates affects generation length: low-rank subspaces alter sequence length without modifying the alignment loss. We present LOCUS, a method that selects a task-aware low-rank adaptation subspace to minimize output-token cost subject to a utility constraint. Within this subspace, post-training retains the native preference objective with a frozen backbone. Across Anthropic HH-RLHF dialogue preferences, we evaluate two $\sim$3B decoder backbones, Pythia-2.8B and Qwen2.5-3B, against protocol-matched full-parameter DPO and DrDPO branches and the released SamPO checkpoint. LOCUS reduces continuation length by up to 39.84% on Pythia-2.8B and by 14.87–17.58% on Qwen2.5-3B while updating only 0.24–0.28% of model parameters, with no material change in the internal preference diagnostic.