Deep research agents synthesize open-web evidence into long reports, but current benchmarks do not evaluate research centered on a source video. We introduce Video-Centered Deep Research Report Generation. Given a video and user request, an agent identifies research questions, retrieves external multimodal evidence, and produces an interactive report linked to relevant video moments. This task has direct practical value because it helps users locate relevant moments, verify claims, reproduce demonstrated procedures, and make informed decisions from video content. We present a skill-based workflow and OmniDR-Bench for this task. For evaluation, we propose Omni Atomic Rubric Trees (Omni-ARTs). Each tree contains three dimensions, 23 shared directions, and task-specific atomic binary criteria. The directions test audiovisual grounding, multimodal evidence use, and interactive presentation. Audited Rubric Construction (ARC) builds these trees through independent evidence collection, automated checks, behavioral calibration, and iterative human review. OmniDR-Bench contains 50 Chinese and English tasks across 19 realworld subcategories. It includes 6,414 criteria. Across 18 agent configurations, the best system scores 63.3 out of 100. Current agents still need broader content coverage, stronger causal and logical analysis across modalities, and more accurate citation support. We further analyze models, harnesses, skill configurations, video categories, and research trajectories. The code and data are released at https://github.com/marinero4972/OmniDR-Bench.

Large Language Models (LLMs) can generate UML class diagrams from natural language requirements, but hallucinations may compromise the correctness and completeness of the resulting diagrams. Existing evaluations mostly assess the final diagram, without examining the intermediate traces of reasoning-capable LLMs, which can reveal where these hallucinations arise. We develop an empirical taxonomy of eleven hallucination sub-types organised into three dimensions: Prompt Faithfulness (deviations between requirements and reasoning), Code Faithfulness (discrepancies between reasoning and generated PlantUML code), and Reasoning Quality (problems within the reasoning). The taxonomy was derived from 60 manually annotated DeepSeek-R1 traces, applied to 120 further DeepSeek-R1 traces through a dual-LLM annotation pipeline, and examined on 20 DeepSeek-V4-Pro traces. Code Faithfulness was the most frequent dimension for DeepSeek-R1 in both datasets, whereas Prompt Faithfulness dominated for DeepSeek-V4-Pro, suggesting model-dependent failure patterns. In the manually annotated traces, reasoning-trace length showed no significant monotonic association with hallucination frequency, whereas requirement complexity was negatively associated with it, particularly for Prompt Faithfulness errors. In a mitigation experiment on 13 Golden UML ModelSet cases with three Qwen3-based models, adding Prompt Faithfulness and Reasoning Quality constraints to the generation prompt raised the pooled mean F1 score by 4.0 percentage points, whereas a post-hoc Code Faithfulness review provided no consistent further benefit. Relationship modelling remained the lowest-scoring element throughout. These results suggest that localising hallucinations within the generation process can inform targeted, prompt-based mitigation. Whether these findings generalise beyond the DeepSeek and Qwen3 model families remains to be tested.

Vector quantization (VQ), in particular lattice-based methods, offer strong rate-distortion trade-offs for low-bit large language models (LLMs), yet their behavior under quantization-aware training (VQ-QAT) remains poorly understood. We scale lattice VQ-QAT to billion-parameter LLMs and multi-billion-token training regimes, and introduce SQuAT, a shape-gain VQ-QAT approach trained with a separate straight-through estimator (STE) for the shape and gain. SQuAT decomposes weights into direction and magnitude: normalized directions are quantized using an $E_8$-derived spherical codebook, while gains are quantized independently with low-bit scalar quantizers. We show that, unlike a standard STE applied to the full quantizer, which ignores the shape-gain structure, these separate STEs induce tangent-space updates for directions, yielding an implicit form of spherical (Riemannian) optimization. To enable scalability, we develop a distributed exact nearest-neighbor assignment scheme, in which workers quantize disjoint weight blocks and communicate only discrete indices, reducing VQ-QAT training time by 12.8$\times$ and allowing us to scale it to 10B tokens. Across instruction-tuned LLaMA-3.2-1B/3B, Qwen3-4B, and base Qwen3-1.7B, SQuAT consistently outperforms standard VQ-QAT STE under identical quantizers and initialization, and achieves SOTA VQ-QAT accuracy.

Activation verbalizers translate the internal representations of large language models into human-readable text. The recently proposed natural language autoencoders (NLAs) learn this by reconstructing activations through a natural language bottleneck. However, it remains unclear how well these methods generalize beyond the setting they were trained for. Here, we study this question for NLAs along four closely related questions: (i) Do NLA verbalizations exhibit compositional structure with respect to their input? (ii) Do NLAs generalize to finetuned versions of the model they were trained for? (iii) Can NLAs recover finetuning objectives in model diffing? (iv) Can NLAs be specialized for model diffing? Across four models from three families, our experiments provide evidence supporting a largely positive answer to all four questions. Specifically, verbalizations of composed analogy activations mostly name the target while rarely mentioning the subtracted word, and an NLA can be reused on lightweight finetunes of its original model without retraining, with reconstruction quality degrading noticeably only for the largest model. Verbalizing activation differences between Taboo model organisms and their original models recovers the concealed secret word for most organisms, and reinforcement learning on activation differences further improves this recovery on held-out organisms. We also find that NLAs can recover the finetuning objective without the finetuned model generating any token, since the most informative token positions precede the user prompt. NLAs do not yet match state-of-the-art model diffing methods on the emergent misalignment detection task, but outperform a blackbox baseline. Our findings suggest that NLAs can meaningfully verbalize differences in activation space, rendering NLAs a promising direction for future work.

While there is increasing evidence that words are stored in the mental lexicon with some phonetic detail, there are open questions about exactly which phonetic details are stored and whether they are functionally relevant for spoken word recognition. Word-level phonetic cues, such as word duration, provide an appropriate test case. Across languages, including Mandarin Chinese, high frequency words are shorter than low frequency words. Using a homophone ordering task, we investigate whether differences in word duration impact word recognition. Mandarin Chinese was examined since it has abundant disyllabic homophones. Our results indicate that participants were able to use word duration to recognize the order of homophone pairs, selected to vary in lexical frequency. Both with high noise levels and when no other cues were available, participants mapped the shorter duration auditory stimulus to the higher frequency word. The results provide evidence that word duration can directly cue the identity of words, suggesting that it may be a phonetic detail incorporated into the lexical representation. We discuss this direct cueing account alongside alternative accounts of how word duration may modulate word recognition.

Silicon sampling generates survey respondents by having a large language model (LLM) answer a survey as a person with given demographic attributes, and the resulting synthetic populations are increasingly used in place of human ones. Existing evaluations focus on individual-answer accuracy, per-question answer distributions, and correlations across questions. None of these compares the respondents themselves, so a generator whose respondents have combinations of answers found in no human respondent can still score well if it reproduces every answer distribution. We propose the respondent coverage gap (RCG), which represents each respondent as a point defined by their complete answer profile and measures how far each human respondent is from the nearest generated one, relative to the distance between human samples. On two US social surveys, all six LLMs we evaluate leave human respondents uncovered, and under our main settings, they fall behind a baseline that samples each answer independently from the human marginals. Across the prompting and inference settings we test, every LLM has a higher RCG than a Gaussian copula fitted to the human data.

Cyberbullying involves behavioral and contextual patterns beyond isolated abusive messages. However, existing approaches remain fragmented across content detection, semantic representation, and network analysis, with limited interpretable perpetrator representations, multi-scale structural modeling, and explanations of behavior-driven similarity and influence propagation. These limitations hinder the identification of persistent perpetrators, characterization of behavioral communities, and assessment of potentially influential users. This paper proposes Ontology-Driven Cyberbullying Behavior and Network Analysis (ODCBNA), a perpetrator-centric framework integrating requirements-driven ontology design, knowledge graph construction, behavioral modeling, community analysis, and influence modeling. Using the Bengali Contextual and Diverse Cyberbullying (BenCoDiC) corpus as a use case, ODCBNA identifies $1{,}063$ users with at least three bullying comments or replies as perpetrators and represents each through a $64$-dimensional behavioral and contextual feature vector. Cosine similarity and $k$-nearest-neighbor modeling produce a weighted network with $27{,}554$ edges. Louvain analysis identifies six communities with a modularity of approximately $0.495$ and $27$ sub-communities. Feature-contribution analysis identifies the Video, Impact, and User Profile groups as the strongest contributors to behavioral similarity and community separation, while contextual dimensions support interpretation. Community-constrained influence analysis shows that IM Greedy achieves small positive gains in all six communities, although different seed-selection methods produce comparable diffusion spreads despite limited seed overlap. These findings provide an interpretable basis for understanding perpetrator behavior, community organization, and potential influence patterns in Bengali online environments.

Large language models increasingly act as controllers in structured decision pipelines, yet their decisions are still executed by autoregressive text decoding. Hierarchical RAG is one such setting: at each frontier of a document tree, an LLM commits to $b$ of $k$ sections under an input-dependent action grammar, executed as either failure-prone free-form JSON or reliable but strictly serial grammar-constrained generation. We show that this serialization is unnecessary for bounded-horizon control: the reachable future forms a shallow action-prefix tree, which we compile --- one root prefill plus one tree-causal probe evaluate every prefix-conditioned action distribution with serial decoding's tokens, positions, and causal ancestry. A prefix-state lemma makes compiled control policy-exact in exact arithmetic: the same policy with a two-forward critical path and zero output tokens. One further probe pass certifies the exact order-marginalized MAP set. Across four document-disjoint splits and two scales on the primary MetaX device, compiled control matches the serial policy on $84$--$89\%$ of questions --- every instrumented divergence a floating-point near-tie --- and cuts control latency by $35$--$47\%$ versus token-wise serial decoding and $6$--$7\%$ versus an idealized chunked engine; an H800 repetition preserves the ordering.

Enthymemes are arguments that leave a premise or conclusion unstated, and they are a common persuasion device on social media. Because readers must fill in the missing part themselves, they often disagree on whether a text contains one. This paper asks how to detect and evaluate enthymemes when such disagreement is part of the task. On tweets about vaccination and immigration, each labeled by five annotators, F1 stalls around 0.67 for every fine-tuned model we train, regardless of data size, added features or label type. We therefore ask how these models should be evaluated instead, and propose metrics that compare predictions with the full distribution of annotator votes, together with a model that predicts this distribution directly and captures disagreement better at no cost in F1. We then ask what role discourse structure plays. Models learn to rely on explicit argumentative structure: they reliably recognize tweets without an enthymeme but rarely commit on tweets that contain one. Both the models and a discourse parser miss arguments whose links are left implicit and must be inferred by the reader. Finally, we discuss what these blind spots mean for content moderation systems trained on majority labels.

Our previous analysis found that a model's native hashed N-gram table is where continual updates disturb the base least, but its writes needed backpropagation and changed the pretrained table. We present Delta N-gram, a forward-only write path into that table. The pretrained table T^0 is never modified. Evidence is written into a sparse patch \Delta on the rows it addresses: a 4.1M-parameter writer emits a target and a gate per (position, hash head) event, and a block-averaged delta rule keeps each row a convex combination of its old value and the gated targets. The writer is trained offline on raw text with the pretraining objective; at use time writing takes forward passes only, and clearing \Delta restores the original model. A 0.6B prototype trained to answer questions keeps 16/16 facts after a second write; online gradient writes keep 2--4/16. We study two writers: a plain writer that encodes evidence with the T^0-only model, and a delta writer that reads its current memory while encoding. On Qwen3.8-Flash-Next, the plain writer's memory is document-specific: on 64 held-out news events it lowers continuation NLL by 0.068 nats/token against empty memory and by 0.024 against another document's memory, and 0.063 and 0.013 after 16 intervening articles. The delta writer's memory is more document-specific and keeps more of it under distraction (70% vs 54%). Over 35,000 writes on our previous seven-task chain both writers stay stable and the delta writer drifts less (layer-2 math drift 176% vs 188%); later accuracy changes by -0.26/-0.52pp (-7.29pp for gradient updates) and small-k_i drift is 17.9/17.5% (39.2%). Acquisition is negative, and general-input drift reaches 32.8% (20.75%): format rows read by every example grow to 4--12 times their pretrained norm. Zeroing eight such rows removes over 98% of layer-2 drift on math inputs, together with their NLL gain.

Wrong automatic routing of an IT support ticket can cost more than human triage. We evaluate cost-aware abstention as an offline, cost-conditional readiness assessment: select a routing policy on calibration data under explicit costs, then evaluate that frozen policy once on held-out test data. On a public 47,837-ticket, eight-class corpus, we compare four trained non-LLM classifiers with four large language models under zero- and few-shot prompting. Under the illustrative cost model (c_ok, c_triage, c_wrong) = (0, 1, 5), TF-IDF with logistic regression and a threshold of 0.60 selected on calibration reduces test cost from 0.763 to 0.472, a 38.1% relative reduction (paired ratio-bootstrap 95% CI [35.7%, 40.4%]), while lowering wrong routes from 15.3% to 3.5%. Frozen calibration-selected policies also beat both always-route and always-defer on held-out test data in 14 of 16 settings in a post-hoc cost-parameter analysis. The strongest prompted configuration reaches 0.477 accuracy on a stratified 1,000-ticket test subset, versus 0.879 for fine-tuned DistilBERT on the same tickets. Isotonic lowers expected calibration error without reliably improving discrimination: GPT-4o-mini few-shot has one confidence value before and after recalibration and correctness AUROC 0.500. We therefore report correctness AUROC and tie-aware risk-coverage area alongside ECE. At the analytically derived break-even probability 0.80, seven prompted configurations defer all tickets; Haiku few-shot routes only 19/1,000. An LLM-only agreement control and margin/entropy rejection baselines further separate calibration from discrimination. The findings are scoped to this dataset, prompt design, model endpoints, and illustrative costs; the work is an evaluation framework, not a new abstention algorithm or universal deployment certification.

Metaphor is a crucial aspect of human cognition. As such, we would like to understand how it is represented in large language models (LLMs) and how it might be controlled. We use mechanistic interpretability techniques to identify causal subspaces in a range of LLMs covering different architectures and sizes. We show that we are able to steer output on classification tasks across different parts of speech. We are also able to steer output on generation tasks, indicating that these techniques can be used to control the metaphoricity of LLM output.

Continual learning requires a model to keep acquiring new skills and knowledge without compromising its existing capabilities. Parametric methods suffer from catastrophic forgetting and loss of plasticity. Non-parametric methods, such as prompts, skills, and harnesses, leverage in-context learning but are limited by context length and lossy plain-text compression. To effectively combine their complementary strengths, we propose HERD (Harness-Evolving Recursive Distillation), a hybrid continual learning framework that learns efficiently in the harness and internalizes experience into cumulative capabilities via distillation. On the harness side, HERD continually refines its skills to capture reusable patterns, while retaining high-residual examples that the skills fail to capture in a bounded memory. On the parameter side, HERD performs drift-constrained distillation with dual anchors, substantially reducing forgetting over long recursive runs. On the 25-stage task stream of TRACE+17, under fixed memory and compute budgets, HERD matches the strongest parametric baseline in new-task acquisition while exhibiting the least forgetting and the strongest retention of general capabilities.

While arousal can be reliably predicted from musical audio, emotional valence remains difficult to model due to its subjective and context-dependent nature. In this work, we investigate the potential of in-context learning (ICL) with large language models to advance continuous valence and arousal prediction. Rather than predicting emotion scores directly, our framework leverages contextual demonstrations from an auxiliary dataset to refine the initial predictions produced by a small, task-specific base model. We systematically evaluate a text-based ICL variant using lyrics and generated audio captions, alongside a multimodal variant that incorporates raw audio signals. Using the entire MER60 dataset strictly as a held-out test set, our experiments demonstrate substantial performance gains, particularly for valence. Text-based ICL improves the valence $R^2$ from 0.110 to 0.427 with 80 demonstration shots, while the multimodal variant achieves 0.346 with only 10 shots. These findings suggest that training-free ICL offers an effective inference-time strategy for current language models to predict continuous music emotion.

Long-context models often use answer-bearing evidence less effectively when it appears away from a prompt boundary. We compare this position bias across Transformer, state-space, and hybrid Mamba architectures by moving one gold document through a fixed set of nine distractors. The evaluated Transformers show a beginning advantage that the evaluated state-space models lack, while both families favor evidence near the end. A matched pure-versus-hybrid Mamba comparison points in the same direction but remains statistically uncertain. The family difference appears among models that can answer the task. Transformer primacy reappears on synthetic retrieval, where state-space ceiling prevents a family comparison. A corpus swap within one state-space architecture changes overall accuracy without a detectable change in curve shape. Prompt changes can create or nearly remove a beginning advantage. Position remains decodable across architectures even when response curves differ, and attention concentration near initial tokens covaries with beginning sensitivity in the tested Transformer family. These exploratory results show that effective context use depends on the evaluated architecture family, capability, task, and prompt, without isolating the sequence mixer as the sole cause.

Content warning: this paper quotes slurs and hateful statements from published test suites; the racial slur among them is masked. Automated systems decide at scale which messages are harmful, and they decide unevenly. Classifiers label African American English as offensive up to twice as often as other English, the accounts of prominent drag queens can score as more toxic than those of white nationalists, and safety behaviour that holds in English collapses in low-resource languages. We argue that over-flagging and under-protection share one cause: deciding harm from vocabulary and surface form rather than from what a message does. We organise the evidence into seven failure modes and argue, as testable hypotheses, that harmful behaviours (a request for secrecy, a demand for payment, an attempt to isolate someone) travel across languages better than the words that express them. We propose an evaluation protocol that makes the gap between the best- and worst-served language communities a headline result, and show that the naive gap is inflated under equal service and that its usual bootstrap interval is unreliable. Applied to five open systems in eleven languages, from a keyword lexicon to two current multilingual guard models, the protocol finds that none reaches native-speaker accuracy at separating harmful function from harmless form (51 of 55 system-language pairs fall short); that greater capability raises discrimination in nine languages while leaving Arabic and romanised Hindi near chance, so the gap between the best- and worst-served languages grows with capability; that one system both over-flags and under-protects the same language most; and that a language a system reads worst can look the most fairly treated when only false positives are reported.

Policy enforcement for LLM agents has shifted from single-turn content moderation to compound, application-specific policies evaluated over multi-turn transcripts and tool-call trajectories. Fine-tuned guardrail models are efficient but degrade on compound policies, and while LLM-as-a-judge approaches achieve higher accuracy, they incur high inference cost and tend to produce unstable predictions. Neurosymbolic methods offer a natural alternative, decomposing the policy into narrow neural decisions composed under explicit logic. However, existing symbolic layers lack the required expressiveness and are optimized against fixed data rather than the policy itself. To address this, we introduce SCOPE, which compiles a policy into an executable program of neural predicates executed by small models, deterministic extractors, and code that composes their outputs. The program is optimized jointly with its training data, using adversarial generation of counterexamples that target where the program deviates from the policy. We evaluate SCOPE on a benchmark of real-world policy tasks against prompt optimization, task decomposition, and workflow generation methods, achieving the highest average accuracy across tasks and models. Programs optimized with the weakest executor transfer to other models without further optimization. This allows them to be deployed in production on inexpensive models, where they reach accuracy comparable to frontier reasoning models. Unlike these monolithic models, SCOPE produces substantially more stable predictions, and each verdict is interpretable through the predicate outputs that produced it. Additionally, we validate the key components of our method through an ablation study. SCOPE thus enables reliable, interpretable, and efficient policy enforcement at production scale.

This paper studies the local geometry of semantic meaning in large language models. Current representation analyses ask whether meanings are close in embedding space, or whether linear probes separate semantic classes, but they do not characterize the within-concept local geometry: what is invariant across realizations of a single semantic factor, what varies, and what changes when meaning is refined. This paper characterizes that geometry through two controlled probes evaluated on four architecturally diverse decoder-only LLMs (Gemma 4, Llama 3.2, Qwen 2.5, Pythia). A cross-lingual probe uses language as a perturbation to separate exact meaning matches, same-factor realizations, nearby concepts, and out-of-domain contrasts. A refinement probe separates specificity-bearing semantic narrowing from high-probability attributive elaboration and adversarial controls; both probes are validated by blind human judgments. Four findings recur across all models: under mean pooling, once dominant variance directions are removed, sentence pairs order by semantic relatedness (same factor, then same domain, then different domain) without overlap over a band of layers; semantic factors organize around an invariant core surrounded by structured realization-specific variation; refinement produces directed displacement into tighter neighborhoods, including against controls built on the same base noun; and sentence-level and token-local readouts expose complementary geometry (volume contraction at the sentence level, sharper anisotropic endpoints at token-local readout). We interpret these results as evidence that LLM semantics are organized as local anisotropic neighborhoods in which meaning contracts under refinement.

The number of attention heads in a transformer is usually set by convention or grid search. We test a simple heuristic for setting it in advance: estimate the intrinsic dimension of token representations without labels, and use roughly that many heads. We estimate intrinsic dimension with a nested-dropout VAE trained on DistilBERT embeddings of WikiText-2, choosing the KL weight at the knee of the VAE's rate-distortion curve rather than by hand. This gives $k^*=9.2$ (95% CI [8.6, 9.8]). In a small 4-layer transformer trained from scratch, 10 heads and 32 heads are statistically equivalent on SST-2 and on MRPC (paired TOST over 10 seeds, $\pm1$ point margin), with 69% fewer attention parameters for the 10-head model. Training 10 heads from the start also slightly outperforms training 32 heads and pruning to 10 (0.6 points, $p=0.008$). The evidence is narrow: two classification tasks, one small architecture, no pretrained models, and no formal link between intrinsic dimension and head count. We present the result as a practical heuristic worth testing at larger scale, not as a general principle.

Language models often produce semantically similar answers to open-ended questions, even when many plausible alternatives exist. Generating useful alternatives requires identifying both what the query calls for and what existing responses already cover. We introduce Gap Detection, a framework that searches for relevant but unoccupied directions in a shared text embedding space. Its objective, GapScore, combines query relevance (Demand) with dissimilarity to the nearest existing item (Defensibility). The framework optimizes this product on the unit sphere, maps candidate directions to concept words with a language model, and uses task-specific filtering and constrained prompting to turn those concepts into text. On five curated topics and 90 INFINITY-CHAT queries, the reported mean relative reductions in output Collision are 12.8% and 10.8%, respectively. We also instantiate the framework for generative engine optimization over four product markets and for iterative synthetic-data augmentation. In the latter setting, with 100 training examples per method, Gap Detection reduces Collision from 0.324 to 0.298 relative to undirected accumulation and increases the Vendi Score from 27.16 to 29.20; the accuracy difference is one example on a 40-example test set. Component ablations expose the trade-off between semantic separation and relevance, while also showing that better direction scores do not uniformly yield better downstream metrics. These results support explicit gap search as a practical mechanism for diversifying content, with task validity checked separately from geometric novelty.