Papers

Multimodal Large Language Models have achieved remarkable progress in short-form audio-video understanding, yet long-form audio-video comprehension remains challenged by limited context windows and severe information redundancy. To address these bottlenecks, we propose AVOC, a framework for long-form audio-video understanding in Omni-modal Large Language Models. AVOC introduces a learnable token compression module between the modality encoders and the LLM backbone. We reframe multimodal token compression as a top-$K$ retrieval problem: given a fixed context budget, the module must retrieve a compact subset of tokens that best supports answering the user query. We draw inspiration from three classical Information Retrieval criteria for selecting informative units from a large candidate pool: relevance, importance, and diversity. AVOC instantiates each criterion as a tailored mechanism for audio-video understanding, and integrates them into a unified retrieval-style compression pipeline. Experiments show that AVOC achieves state-of-the-art performance on long-form audio-video benchmarks, surpassing the second-best model by 4.9 and 5.5 points in average accuracy on OmniVideoBench and LVOmniBench, respectively. Moreover, AVOC maintains robust performance on Audio-Video Needle-in-a-Haystack task at durations up to one hour. Code and model are at github.com/YJCX330/AVOC.

Reasoning language models are increasingly asked not only to answer difficult questions, but also to estimate their likelihood of success. Existing methods typically elicit confidence only once: either before thinking or after answering. We argue that confidence in reasoning models is state-dependent: before thinking, confidence should estimate the chance of the model correctly solving the prompt, while after thinking it should predict whether the realized answer is likely to be correct. This distinction determines the appropriate supervision target: prompt-level success should supervise confidence estimates made after seeing the prompt, while individual answer-level correctness should supervise confidence estimates made after answering. We introduce CALIBER (Calibration Before and After Reasoning), which elicits both estimates and supervises each with the target matched to its information state. Under this unified protocol, CALIBER reduces Expected Calibration Error (ECE) by 52.5% over the strongest single-confidence baseline on BigMathDigits for the 7B model, while achieving the best Brier score and AUROC, and remains within 2.1 points of the best accuracy. Further, on a larger 30B model, CALIBER achieves the best ECE on BigMathDigits while remaining competitive in Brier score and AUROC. Out of distribution, it achieves the best ECE and Brier score on GPQA and TriviaQA, and remains competitive on SimpleQA. Ablations further show that this position-target alignment is most beneficial under distribution shift where it consistently reduces calibration error across all out-of-distribution benchmarks.

While in-context learning is generally shown to be effective in Large Language Models (LLMs), bad contexts can cause performance degradation and mode collapse, a phenomenon we call "pigeonholing." **Unintentionally bad** contexts can happen without malicious jailbreaking intents: For example, a user asks the model to justify an incorrect math theorem or fails to correct the model's buggy code. Specifically, we investigate “pigeonholing" in two scenarios: (1) when the user suggests a solution, and (2) when the conversation context includes the assistant's previous (incorrect) responses. Our experiments across 10 verifiable and open-ended tasks with 10 different models show that pigeonholing manifests in several ways: (1) repeating the incorrect answers from context (leading to 38-40% performance drop), (2) converging on a narrow set of answers in coding and text generation without exploring alternatives, and (3) flipping stance on controversial topics to align with the user or the assistant's previous claims. We find that pigeonholing worsens almost monotonically with the number of conversation turns (performance drops by additional 14+% as repeated mistakes increase from 1 to 5), and pigeonholing-induced mode collapse can happen even when the provided example is correct. As a step toward mitigation, we propose RLVR with synthetic errors which improves models by 43-60% under bad contexts compared to vanilla RLVR baselines.

Fine-tuned encoders deployed across heterogeneous NLP tasks face three compounding problems: mismatched inductive biases, class-imbalance corruption of feature statistics, and no mechanism to condition attention on external lexical knowledge. We introduce \surgellm, a unified transformer framework that addresses each with a dedicated lightweight module: a surgical feature gate (learned per-dimension sigmoid over curated lexical indicators and [CLS]; provably degenerates to identity when features are uninformative), task-conditioned prefix tokens (quantized feature values and task identity prepended to every input), and Instance-Weighted Normalization (IWN; removes class-prior bias from gate statistics). We prove an excess-risk bound linking gate benefit to surgical feature alignment. Across four tasks, SST-2, multi-hop retrieval, LLM-prompt attribution, and authorship detection, covering 17,830 examples and eleven model variants over three seeds, the IWN variant achieves macro-F1 0.940 ($+0.036$ over the strongest non-IWN baseline; $+0.130$ on authorship detection). A random-vocabulary control ($-0.028$ avg.\ F1) confirms gains are lexical, not parametric. Code, vocabularies, and a $99.5\%$-recovery auto-extraction recipe are released.

Stochastic quantum neural networks (SQNNs) encode neuronal activations as qubits, synaptic topology as entanglement, and neural noise through a Lindblad master equation. A recent conference study applied a ring-entangled SQNN to collaborative intrusion detection and reached three conclusions: ring entanglement is essential for non-local anomaly detection; an adversarial-resilience bound holds but is conservative; and the depolarising channel fails to act as a dropout-style regulariser, behaving instead as output noise. It left open whether a per-gate stochastic deactivation (“true quantum dropout”) could regularise where the depolarising channel could not, and whether the loose robustness bound could be replaced by a predictive theory. This paper resolves both and extends the framework to real data and to neutral-atom hardware. We give an $N$-qubit formulation through the stochastic master equation and its vectorised Liouvillian, and prove a decoherence-contraction theorem: a depolarising channel of strength $\gamma$ over $L$ entangling layers contracts every weight-$w$ Pauli read-out by a factor $(1-4\gamma/3)^{wL}$ (for the weight-$1$ read-out used here, $(1-4\gamma/3)^{L}$); building on the general noise-as-defence result of Du et al., we make this quantitative and operational for intrusion detection. On the real NSL-KDD dataset under white-box FGSM and PGD attacks, a depolarising SQNN trained with the channel is, over seven seeds under strong $\ell_\infty$/$\ell_2$ attacks, significantly more robust than the noiseless circuit ($\ell_\infty$ PGD-$20$, $p=0.04$, large effect) and, critically, never suffers the catastrophic robustness collapse that the noiseless model and gradient-trained classical detectors (which fall from $95\%$ to $47\%$) do, cutting robustness variance roughly twofold; we show this robustness arises from a noise-reshaped training boundary rather than from attack-time gradient contraction. For generalisation, we derive an adaptive-penalty formula showing that per-gate dropout implements a curvature-weighted $L_2$ penalty $\tfrac{p(1-p)}{2}\sum\theta^2\partial^2_\theta L$ in weight space, maximised at $p=1/2$, whereas depolarising noise implements an output-space penalty. A $30$-seed study confirms the formula's quantitative prediction: both mechanisms reduce the train-test gap by a small but statistically significant margin ($\approx\!0.01$; $p<10^{-4}$ and $p=0.004$), are statistically indistinguishable from each other, and the effect is concentrated where overfitting is largest; increasing the dropout rate past $1/2$ does not help, as the formula predicts. The single-seed dichotomy of prior work does not survive replication. We close with a neutral-atom realisation and a feasibility-by-$N$ analysis.

Retrieval-augmented generation (RAG) in clinical settings increasingly requires multilingual retrieval against predominantly English evidence corpora. Multilingual medical retrieval demands three capabilities: cross-lingual alignment, concept discrimination, and evidence retrieval. However, existing benchmarks evaluate these only in isolation, leaving the interaction between biomedical expertise and multilingual coverage unmeasured. We introduce MMed-Bench-IR, a benchmark designed to disentangle these axes across 6 languages and three structurally heterogeneous tasks: (1) cross-lingual medical QA retrieval with 6,127 queries grounded in the Unified Medical Language System (UMLS), (2) concept discrimination over 4,975 confusion sets at three difficulty tiers, and (3) multilingual evidence retrieval for RAG with 2,040 quality-assured queries. The three tasks share zero concept and query overlap by design, ensuring that aggregate scores reflect genuine capability breadth. Evaluation of ten systems across six paradigm families reveals severe cross-lingual failure: biomedical encoders that score 0.818 nDCG@10 in English drop to 0.056 in Japanese, a gap that English-only benchmarks cannot detect.

Conversational product search assistants offer a more expressive, natural, and interactive alternative to traditional keyword-based product search. With limited screen space, showing only a few items increases the need for precise preference elicitation, which can prolong conversations, leading to user frustration and session abandonment. Conversely, rushing to recommend items without a clear understanding of preferences risks poor matches and a degraded user experience. We present Dialogue to Discovery (D2D), an attribute-oriented preference elicitation framework that dynamically exploits the structure of product attributes to efficiently steer conversations toward the user's desired item. D2D adaptively prioritizes the most informative queries and strategically times product recommendations, reducing premature or off-target suggestions that harm engagement. To evaluate D2D, we curate three datasets from the Amazon Reviews corpus. In simulated conversations modelled using a multi-factor utilitarian patience framework, D2D achieves a 22.2-29.9% improvement in target-finding accuracy, 6.6-16.1% reduction in abandonment, and 27.5% shorter average conversations over the state-of-the-art baselines. A complementary user study further confirms significant gains in both user satisfaction and perceived efficiency.

Unlearning has emerged as a key technique to mitigate harmful content generation in diffusion models. However, existing methods often remove not only the target concept, but also benign co-occurring concepts. As illustrated in Fig.1, unlearning nudity can unintentionally suppress the concept of person, preventing a model from generating images with person. We define these undesirably suppressed co-occurring concepts that must be preserved CARE (Co-occurring Associated REtained concepts). Then, we introduce the CARE score, a general metric that directly quantifies their preservation across unlearning tasks. With this foundation, we propose ReCARE (Robust erasure for CARE), a framework that explicitly safeguards CARE while erasing only the target concept. ReCARE automatically constructs the CARE-set, a curated vocabulary of benign co-occurring tokens extracted from target images, and leverages this vocabulary during training for stable unlearning. Extensive experiments across various target concepts (Nudity, Van Gogh style, and Tench object) demonstrate that ReCARE achieves overall state-of-the-art performance in balancing robust concept erasure, overall utility, and CARE preservation.

Mining sentiment information from the textual content of peer review comments offers valuable insights into the scientific evaluation process. However, previous studies are often constrained by coarse-grained analysis and the lack of differentiation across review rounds. Notably, the dynamic shifts in reviewers' focus and sentiment tendencies throughout multiple review stages remain underexplored. To address this gap, the present study investigates the distribution and evolution of aspect-level sentiments and examines their correlation with the number of review rounds. We begin by segmenting the multi-round review comments of 11,063 accepted papers from Nature Communications and identifying fine-grained review aspect clusters. A manually annotated corpus of approximately 5,000 review sentences is then constructed. Using this dataset, we train a series of deep learning-based aspect sentiment classification models. Among them, the LCF-BERT-CDM model achieves the best performance, with a Macro-F1 score of 82.65%. Subsequent statistical analysis reveals a consistent trend: as the number of review rounds increases, the proportion of positive sentiments rises, while negative sentiments decline. Correlation analysis further indicates that aspect sentiment scores are negatively associated with the total number of review rounds. Key aspects exhibiting stronger correlations include "experiments", "research significance" and "result analysis".

Large language models are making research production scalable, shifting the bottleneck from producing artifacts to judging claims. We present Agon, a research orchestrator that validates what can be checked inside the workflow and leaves the remaining judgments to human scientists. Agon is built on six design principles: Prompt Economy, Future-Facing, Minimal Prompts, OmniDisciplinary, Massive Parallelism, and Zero-Code. We ran Agon across domains for 444 iterations of Prompt Economy loops, using only small starting topics and no human-written experimental code. These deployments demonstrate scalability while exposing new classes of failure. We organize these failures into a taxonomy along severity, fixability, visibility, and capability locus. The taxonomy separates failures the loops can see and fix from those that require human judgment. Together, these results show that Agon is pushing research toward a new paradigm: machine scales, human steers.

Reliable structural health monitoring (SHM) of offshore wind turbine (OWT) support structures requires fast state estimation from sparse measurements. Repeated high fidelity finite element or aeroelastic analyses are difficult to use directly in online monitoring loops, while purely data-driven surrogates can require large training sets. This paper presents Digi Turbine, a synthetic reliability-aware Physics Informed Neural Network (PINN) benchmark for OWT monopile support structure monitoring. The workflow embeds a simplified Euler Bernoulli beam equation with Winkler soil foundation in the training objective, couples it with Bayesian-prior-informed inverse identification, and adds First Order Reliability Method (FORM) screening. All validation uses synthetic configurations with analytical or finite-difference ground truth motivated by the NREL 5MW reference turbine context.

More than a billion people communicate in Indic languages, yet the natural language processing infrastructure serving them remains fragmented and underdeveloped. The cause is structural: the field organizes its tools and benchmarks around individual languages or small subsets of genealogical language families, building separate analyzers, parsers, and datasets for each language and starting over for the next. This overlooks a deep regularity. Through more than two millennia of convergence around Sanskrit, Indic languages came to share a morphosyntactic architecture formalized in Pānini's grammar, the Astādhyāyī. This cuts across genealogical lines, uniting languages through a common framework. We argue that this Pāninian framework supplies a unifying computational architecture the field has lacked, and that benchmarks grounded explicitly in it would make Indic language systems more accurate, more data-efficient, and more transferable, effectively merging many apparently disparate and sparse Indic language resources into a single high-resource metalanguage bedrock. We propose a four-part benchmark suite to render this shared architecture explicit, measurable, and ready to be leveraged for practical applications. Moreover, we underscore the question it raises for interpretability research: whether neural models trained on these languages come to represent Pānini's categories on their own.

Reliable provenance for LLM outputs requires multi-bit watermarks that remain robust under editing while maintaining low false-positive rates. Existing ECC-based LLM watermarks rely on hard-decision decoding, discarding token-level reliability information and limiting robustness under post-generation edits. We propose CORE-BREW, a COnstant-hit-Rate Embedding extension of BREW for multi-bit watermarking. CORE-BREW calibrates the watermark channel by targeting a fixed hit rate $p^\star$, yielding closed-form per-token log-likelihood ratios (LLRs) for soft-decision decoding. It incorporates entropy-aware erasures to limit perturbations in low-entropy contexts and combines likelihood-based scoring with soft-decision list decoding to exploit soft evidence. Experiments on open-source LLMs under token-level edits and paraphrasing demonstrate that CORE-BREW generally improves detection robustness and payload recovery over the BREW baseline while maintaining low observed false-positive rates. Despite higher conditional perplexity, BLEU and BERTScore remain close to those of unwatermarked text, indicating comparable reference-based translation quality.

Foundation models have been increasingly applied to behavioral science domains such as psychology, sociology, and economics. While these models show promise in tasks such as survey response prediction and human-subject experiment simulation, there remains no systematic understanding of how well they perform across diverse behavioral science tasks. We introduce BehaviorBench, a comprehensive benchmark that evaluates foundation models along four core capabilities: (1) behavior prediction and simulation, (2) strategic decision-making, (3) subject-trait inference, and (4) behavioral knowledge application. Crucially, BehaviorBench evaluates model outputs at both the individual and distributional levels, capturing not only per-subject accuracy but also population-level alignment, an essential requirement for behavioral validity. Our evaluation shows that BehaviorBench remains challenging for leading general-purpose LLMs and behavior foundation models that are specifically trained with behavioral data. We find that individual-level and distributional performance do not always align. General-purpose LLMs tend to underestimate the diversity of human responses, whereas behavior foundation models often lag behind at individual-level prediction. Our investigation further demonstrates how fine-tuning on diverse behavioral data can improve both individual-level prediction and distributional alignment, balancing these two objectives. Our results highlight the importance of evaluation at both individual and distributional levels, establishing BehaviorBench as a foundation for developing and assessing behaviorally aligned AI systems. Our BehaviorBench and models can be accessed via https://umich-foreseer.github.io/behaviorbench/

Existing medical AI benchmarks lack process visibility, atomic skill evaluation, and integrated hallucination detection. We introduce MedBench v5, a redesigned benchmark for clinical multimodal models (language, vision-language, and agent systems) that moves from static QA to dynamic, process-oriented evaluation. MedBench v5 features: (1) a dual-dimensional framework combining Clinical Cognitive Responsiveness (13 sub-dimensions) and Medical Atomic Skills (4 agent environments), covering 63 tasks; (2) three switchable information-flow stressors (omission, contradiction, evidence delay) for factorized degradation analysis; (3) a dynamic process audit protocol with five reasoning nodes that produces model-specific failure fingerprints; (4) hallucination propagation monitoring across initiation, propagation, anchoring, and contradiction interaction-capturing silent hallucination. Experiments on frontier models show that strong overall task performance does not guarantee process stability: stressors mainly disrupt contradiction detection, diagnosis updating, hallucination propagation, and contradiction-based self-correction, while final evidence grounding can remain superficially stable. MedBench v5 provides a unified infrastructure for capability profiling, controllable stress testing, process auditing, and hallucination trajectory analysis in clinical AI evaluation.

Self-evolving agents improve over time by distilling experience from past executions and reusing it in future tasks. Existing systems represent such experience either as natural-language text injected into the agent context or as code exposed as callable tools. However, the choice between these representations is typically made at design time rather than derived from the characteristics of the experience itself, leaving the trade-offs between them poorly understood. We present the first controlled study that isolates text memory and code memory over an identical set of experiences. Our results show that the two forms exhibit complementary trade-offs in construction cost, execution efficiency, and transferability, such that neither representation alone is sufficient. Guided by these findings, we propose Metis, a self-evolving agent system built on a hierarchical dual-representation memory. Metis organizes textual experience into execution plans, environment facts, and common pitfalls, and selectively crystallizes recurring plans into validated callable tools. This design combines the broad applicability of text memory with the execution efficiency of code memory while incurring tool-generation cost only when justified by repeated reuse. We evaluate Metis on AppWorld, a challenging benchmark for interactive agents. The results show that Metis improves task accuracy by up to 20.6% over ReAct while reducing execution cost by up to 22.8%. Compared with representative self-evolving agent systems, Metis consistently achieves a better balance between accuracy, execution efficiency, and memory-construction cost.

Aligner-Encoders are recently proposed seq2seq end-to-end ASR models that replace decoder attention by predicting the uth token directly from the u-th encoder position, so the encoder must learn the alignment internally without cross-attention or a transducer lattice. In practice, this alignment often forms abruptly in the upper layers, making training sensitive and brittle on long utterances. We propose InterAligner, which adds an intermediate Aligner objective so alignment can form progressively across depth, together with an intermediate CTC loss (InterCTC) to stabilize optimization. On LibriSpeech with a 17-layer Conformer, a final-only Aligner reaches 5.0/7.8 WER (test-clean/other). InterCTC improves to 3.4/6.0, and InterAligner further reduces WER to 3.1/5.6 with the largest gains on long utterances.

The composition of training data, governed by the diversity of sources and their mixing strategy, is a cornerstone of Large Language Model (LLM) pre-training. Online Data Mixing (ODM), the technique of adaptively adjusting data mixtures during training, has emerged as a promising direction to improve efficiency. However, existing methods are constrained by their reliance on a singular optimization perspective, which fundamentally overlooks the need for complex LLM pre-training to consider the dynamic data composition from multiple dimensions. To overcome this limitation, we introduce the Holistic Data Scheduler (HDS), a novel online data mixing framework. HDS formulates the data scheduling challenge as a reinforcement learning problem in a continuous control space and leverages the Soft Actor-Critic (SAC) algorithm for its stability and sample efficiency in exploring the high-dimensional policy space. At the core of HDS lies a novel multi-objective, holistic reward function that integrates three critical perspectives: a data-driven reward for quality, a loss-driven reward capturing inter-domain influence, and a model-driven reward based on weight norms. To validate our design and determine its optimal configuration, we conducted systematic experiments on LLMs of various sizes. On The Pile benchmark, HDS reaches the final validation perplexity of the next best method with 44% fewer training iterations. Furthermore, it achieves a 7.2% improvement on the MMLU 0-shot task along with consistent gains on other benchmarks, showcasing its ability to enhance both training efficiency and final model capability.

Discrete diffusion language model (DLM) fine-tuning inherits inexpensive diagnostics from denoising-time confidence monitors, but their PEFT-training meaning is untested. We test top-1 argmax concentration as a collapse warning. Across 816 LoRA/PEFT configurations from three DLM families, the warning fires for every configuration while logs record 0/816 actual collapses at the 200 step horizon, giving zero precision. The cause is pre-equilibrium saturation: top-1 concentration is already high before optimization and quickly becomes insensitive to final training stability. We then evaluate max LoRA gradient norm, a parameter-side signal that samples gradient routing rather than token concentration. On a pooled held-out LLaDA-family split, a train-optimized threshold identifies top-decile final-loss configurations with precision 0.68 and F1=0.79, above the all-positive top-1 baseline even at the lower split-bootstrap confidence bound. Autoregressive controls and cross-family threshold failures bound the result to short-horizon DLM-LoRA inspection rather than a universal collapse detector. Workflow: drop top-1 as a PEFT alarm, log max-gradient early in training, and calibrate thresholds per DLM family before routing runs for inspection.

Most electronic health record (EHR) foundation models encode clinical events as discrete event tokens from a fixed vocabulary and therefore cannot directly represent events containing unseen concepts or new combinations of concepts and attributes such as numeric values. This limits transfer across institutions and even across deployment pipelines within the same institution. We introduce PORTER, a language-grounded structured EHR foundation model that decouples event representation from this fixed vocabulary. PORTER represents events through their descriptions using a frozen text encoder, integrates numeric values through a dedicated pathway, and learns clinical dynamics over patient timelines with an autoregressively pretrained temporal backbone. Across 74 clinical prediction tasks at a pediatric hospital, PORTER matched the mean AUROC of a fixed-vocabulary model with the same temporal backbone and pretraining objective. When the same patient timelines were rendered using event descriptions not seen during pretraining, PORTER transferred without retraining or vocabulary mapping, recovering 97.1% of the mean AUROC of a model trained directly on the target vocabulary. When transferred to MIMIC, PORTER outperformed the fixed-vocabulary model, which dropped 69% of events because their tokens were unseen. Mechanistic analyses showed cross-vocabulary transfer tracked preservation of patient-level representation geometry rather than the scale of the text encoder, and the numeric pathway improved sensitivity to magnitude without disrupting clinical concept identity. PORTER also achieved higher AUROC than a task-specific text serialization comparator, at 329-fold lower amortized compute. PORTER is a step toward vocabulary-independent EHR foundation models that reduce the need for vocabulary harmonization while preserving in-domain performance and enabling efficient cross-task reuse.

Algorithms have become central to scientific research in the era of artificial intelligence (AI). Although algorithm mentions in papers are often used to indicate popularity and influence, existing studies usually evaluate individual algorithms in isolation and pay limited attention to the collective influence formed through their interconnections. This study constructs large-scale algorithm co-occurrence networks in natural language processing (NLP) based on the full text of academic papers and investigates algorithm influence from a network perspective. Using deep learning models, we extract algorithm entities and build overall, cumulative, and annual co-occurrence networks. We analyze their structural characteristics and apply multiple centrality measures to assess the group influence of algorithms across the whole field and over time. The results show that algorithm networks display typical features of complex networks, with increasingly dense connections developing over approximately two decades. Classic, high-performing algorithms and those located at the intersections of different research periods tend to have high popularity, control, centrality, and balanced influence. When the influence of an algorithm declines, it usually loses its core network position first, followed by weaker associations with other algorithms. This study is the first large-scale analysis of algorithm co-occurrence networks. Covering more than four decades of academic publications, it provides a temporal and structural view of algorithm influence and offers a foundation for future research on networks linking algorithms, scholars, and tasks.

We ask whether the geographic origin of Tang-dynasty poets leaves a detectable linguistic trace in their work. Aggregating every poem attributed to each author in the Complete Tang Poems (Quan Tang Shi) and linking poets to their administrative circuit of origin via the China Biographical Database (CBDB), we build a poet-level corpus of 357 poets across the ten Tang circuits and frame origin prediction as multi-class classification. Using character $n$-gram TF-IDF together with interpretable domain features (imagery, season, and allusion), classical and neural models predict a poet's broad region (South vs.\ North) at $0.69$ accuracy, well above the $0.53$ majority baseline, and finer circuit-level origin above chance. Beyond classification, three findings emerge. (i) Linguistic distance between circuits grows with geographic distance (Mantel $r=0.40$, $p\approx0.09$ over nine circuits), evidence of a distance-decay effect in poetic language. (ii) The signal interacts with time: South/North separability is at chance in the High Tang and strongest in the Late Tang, consistent with court-driven homogenization at the empire's height followed by regional divergence. (iii) The model's confident errors are historically meaningful – in the Early Tang, every misclassification is a southern poet read as northern, reflecting the prestige of the northern court idiom. We further show that, when given the whole corpus through a hierarchical frozen-encoder representation, a classical-Chinese transformer (GuwenBERT) only matches – not beats – simple TF-IDF, and that combining them adds nothing, indicating that character $n$-grams already capture the regional signal. Our results position interpretable machine learning as a hypothesis generator for literary history.

On-policy distillation (OPD) trains a student policy using teacher signals computed on trajectories sampled by the student itself. Recent work shows that sampled-token OPD can be fragile on long-horizon reasoning tasks and that local teacher-support matching is a simple and effective repair. This paper introduces blockwise policy-drift gating, a lightweight student-only old-current drift controller for OPD under rollout reuse. The method computes log-probability shifts between the behavior student and the current student on the sampled token path, aggregates these shifts over fixed blocks or spans, and uses the resulting detached, mean-normalized gates to reweight OPD position losses. It does not change teacher targets, teacher top-K supports, or the rollout policy. In a six-variant Qwen3 math reasoning benchmark with a uniform 200-step training budget for all trained variants, we use pass@8 as the primary problem-level solve-rate metric. Fixed 64-token block gating improves sampled-token OPD mean pass@8 from 0.4978 to 0.5160 across AIME24, AIME25, MATH500, and AMC23. On Teacher-TopK/LSM, Block64 gives the best four-benchmark mean pass@8 among trained students. The results identify local old-current policy drift as a practical control signal for reused OPD rollouts and motivate block-level gating as a simple default for improving solve-rate robustness.

"Talk short. Drop grammar. Save token." This caveman style is widely promoted as a way to cut inference cost, but whether it actually saves anything depends on which channel (the user's prompt or the model's response) is being compressed. We present Cavewoman, a two-channel evaluation protocol that scores every generation on task accuracy, realized per-item cost, and reference-text agreement against the model's unconstrained reference. We evaluate eight models on five datasets at five reduction levels, with both channels measured on the same items. Output compression cuts realized cost on most API models (1.4-2.4x per model, up to 3x in the best case) and on all four open-weight models under public-tier pricing. Input compression has the opposite effect, a strict lose-lose: it raises net cost rather than lowering it (~1.15x on the five-benchmark mean, up to 1.8x on the worst dataset and 2.7x under stronger compression), because models compensate with longer responses even as accuracy collapses. Under the same setting, surface text diverges from the unconstrained reference: on the non-reasoning models, roughly half of all generations are correct yet their surface text no longer entails the model's own unconstrained baseline generation. The divergence survives length-controlled re-scoring, multiple-comparisons correction, and replication under complementary semantic measures. Code and data are available at https://github.com/danielle34/cavewoman.

Contextual entrainment, which is a newly discovered phenomenon in large language models (LLMs), refers to the tendency of a model to assign higher probabilities to tokens that appear in its context. In this work, we extend this phenomenon from the token level to the sentence level by examining the per-token mean log-probability of a sentence instead of the probabilities of individual tokens. We investigate sentence-level contextual entrainment across 26 LLMs from seven families and two datasets, which cover both subjective and objective tasks. We find that sentence-level contextual entrainment exists. This means that the sentences in the prompt (even if they are counterfactual statements) can significantly increase their probability during model inference time. As the model size increases, contextual entrainment gradually decreases. We also find that contextual entrainment is controlled by 2% to 4% of the attention heads. Turning off these attention heads can effectively mitigate contextual entrainment without hurting the model's performance.

Speaker recognition has advanced rapidly with large-scale training datasets, yet Vietnamese remains under-resourced, with existing corpora limited in scale and acoustic diversity. Most large-scale datasets rely on facial cues to link speech with speaker identities, restricting data collection to recordings where speakers appear on camera. We propose a face-independent dataset construction pipeline and introduce VieSpeaker, a large-scale Vietnamese speaker recognition dataset. Our approach leverages textual metadata and large language model reasoning to infer speaker identities from transcripts and contextual information. VieSpeaker contains approximately 902 hours of speech from 4,715 speakers. Experiments show that models trained on VieSpeaker achieve improved robustness and generalization compared to existing Vietnamese datasets. This work demonstrates the feasibility of face-independent dataset construction and provides a new direction for building large-scale speech resources.

Modern spoken language understanding (SLU) systems are increasingly deployed in real-world settings, where specific functionalities may need to be removed due to policy or safety constraints. In SLU, a functionality corresponds to an intent and its associated slot-generation behavior. However, in autoregressive models, suppressing a target intent does not eliminate the conditional mapping that generates slots conditioned on that intent. When the intent prefix is externally supplied, the model can reconstruct the original intent-slot structure. We identify this structural failure as \textbf{capability persistence}. We propose \textit{\underline{B}inding \underline{S}ubspace (BSU)}, a representation-level framework that isolates and attenuates intent-conditioned directions underlying this mapping. Across SLU benchmarks, BSU substantially reduces forced-prefix recoverability while preserving retained performance.

Sentiment analysis in Twitter datasets is important because it enables monitoring public opinion on products and analysis of political and social movements. One critical step is preprocessing: the automated processing of text for machine learning algorithms. Preprocessing plays a critical role in reducing noise and improving efficiency. However, little research has systematically examined the order in which preprocessing techniques are implemented. We find that, when accounting for order, spelling correction is the least impactful preprocessing technique, whereas tokenisation is the most impactful. Stemming and stop-word removal are interchangeable, and it is better to remove stop words without removing negation. The best order for applying the preprocessing techniques was tokenisation, text cleaning, stemming, and then stopword removal. Our results provide a systematic approach for practitioners to deploy preprocessing to improve model output without the costly preprocessing exploratory phase.

MeMo proposes language models with explicit multi-layer correlation matrix memories (CMMs), where memorization, retrieval, and forgetting are architectural operations. This paper asks how such memories can reduce the need for retraining when knowledge changes. For changes expressible as MeMo memory associations, the model's accessible knowledge can be updated by editing explicit memories rather than retraining the whole model. We propose a version-aware operation layer in which high-level operations such as replace, obsolete, keep-history, rollback, and trace are compiled into MeMo-native primitive calls over sequences and tokens. The key observation is that a version-aware operation is rarely a single MeMo association. It is an ordered transaction of primitive edits, for example forgetting one sequence-token chain, memorizing another, preserving a historical chain, and recording an inverse program. The framework introduces two auxiliary CMMs: a Version CMM (V-CMM) for mapping version transitions to transaction handles, and a Transaction CMM (T-CMM) for storing reusable change contents and inverse programs. It supports both direct sequence-level edits and structured diff-level inputs, and outlines an evaluation route for update success, rollback, traceability, locality, and transaction reuse.

Existing low-bit KV-cache quantizers often treat each cached key as a flat vector. Under RoPE, however, a key's contribution to a future attention logit decomposes into a position-dependent sum over two-dimensional frequency blocks. This makes key-cache quantization a block-wise bit-allocation problem: high-energy RoPE blocks are more sensitive to quantization error and should receive more bits. We introduce Block-GTQ, a RoPE-aware bit allocator for key-cache quantization built on TurboQuant-MSE(TQ-MSE). For each layer and KV head, Block-GTQ computes a label-free energy score for each RoPE block and greedily allocates integer bit widths by marginal gain. Under matched K/V bit budgets, Block-GTQ better preserves RoPE query-key logits on a ten-model diagnostic panel, cutting per-layer MAE by 32-80% at 2 and 3 b/dim K-only quantization and winning all 367/367 layer comparisons against uniform TQ-MSE. These fidelity gains translate to stronger downstream long-context retrieval, understanding, and reasoning. At K2V2 on Llama-3.1-8B-Instruct, Block-GTQ raises the six-task NIAH average from 70.6 to 97.4, and the LongBench-EN average from 36.87 to 53.31. On AIME 2024/2025 with DeepSeek-R1-Distill-Qwen-7B, without an fp16 recent-key buffer, Block-GTQ at K3V2 scores 51.7/37.5, close to fp16's 54.2/37.9, whereas uniform TQ-MSE collapses to 0.0/0.0. We further implement a packed-cache serving path. On a single H800 GPU with Qwen2.5-3B-Instruct, packed K3V3 achieves 3.24x KV-cache compression with fp16-comparable quality, runs 1.34x faster than fp16 FlashAttention2 at 128K context, reduces peak memory from 56.31 GB to 19.85 GB, and remains feasible at 256K and 512K where fp16 OOMs. Code is available at https://github.com/JIA-Lab-research/blockgtq.