Papers

Black-box model identification works by scoring a model's response to natural-language prompts. One line of work feeds models a degenerate input – their own token, repeated – to find a failure mode rather than an identity. We take that input and ask where the model goes when it does not. For each token t, read argmax p(. | t, t) in one forward pass; the result is a map on the whole vocabulary, with two halves. The first – which tokens are fixed points – is partially anticipated, and we report it as a failed estimand: the natural distance on it is 83% cardinality, separates a corpus manipulation by two bits in 3471 against a precision floor of zero, and attributes families at 0.5833. The second half, where the map sends tokens that are not fixed points, is unrecorded; the one paper holding those tokens logged them as a zero. Pairing on the source token removes the cardinality confound by construction (r from 0.9128 to -0.0932) and attributes families at 0.8333 – twelve models scored against a pool of nineteen – with chance 0.1389, across seven tokenizer groups and several corpora. Two nulls clear it: frequency-matched destinations agree at 0.1429, independent marginals at 0.0798. Family predicts agreement better than tokenizer (0.2031 against 0.1205), and recurrent architectures cluster at balanced accuracy 1.0 against a 0.7895 majority rate, or 0.90 once each model's dominant destination is excluded – the figure we stand behind. We measure the robustness envelope: 8-bit weight rounding moves the map less than deduplicating the training corpus does (0.9004 against 0.6353, on one support), 4-bit destroys it (0.0098; 0.1812 at deployment granularity, so not a coarseness artefact), and the precision floor varies by model from 0.201 to 0.9778. All estimands and kill conditions were registered before the data, and the failed one is reported as fully as the surviving one.

Multimodal Machine Translation aims to incorporate additional signal from non-textual modalities to improve translations by resolving ambiguities. While models, through multimodal fusion, are able to accept images related to the source text, they can ignore this information. Therefore, increasing their visual sensitivity remains an active research area. In this work, we introduce a training method, Metric-based Loss Weighting, that improves visual grounding of translations by increasing the loss function for tokens that benefit from the accompanying image. We identify these tokens using the Point-wise Cross-mutual Information (PCXMI) metric, which compares the model's output probabilities with and without visual context. We introduce a Congruency-based PCXMI metric and experimentally show that both metrics working in combination yield the best results. We evaluate our method by fine-tuning three pretrained Multimodal Large Language Models on the task of Image-guided Machine Translation for three language directions. Metric-based Loss Weighting outperforms other tested methods on the CoMMuTE contrastive dataset, improving accuracy by up to more than 7 percentage points compared to standard fine-tuning, while maintaining strong general translation performance.

Models trained with reinforcement learning for calibrated decisions (RLCD), such as Jev, answer a typed question about an input, the state, with a probability, a choice, or a score, and software acts on the answer without a person reading it. Their robustness has not been measured: adversarial benchmarks score what a model generates or executes, whereas a typed model generates nothing and returns a well-formed answer even when manipulated. Measurement is also hard, because identical requests can return different answers, most available labels come from the model itself, and the API preprocesses each request out of view. Our key idea is to score each attacked decision against the model's own clean decision rather than against labels, and to read it against the change caused by an identical re-run. Building on this, we introduce JevAdvBench, to our knowledge the first adversarial benchmark for RLCD models, with 812 typed questions over 66 scenarios, and a black-box attack suite of 9,744 single-edit variants that each edit one part of a request, with billed input tokens confirming that the edit reached the model. On jev-1.13.0, rewording stays within 1.2 percentage points of the re-run baseline, and fields outside the schema never reach the model. In contrast, one unverified opinion appended to the state flips 12.1% of decisions, statistically tied with the strongest injected command (10.1%), and pushes 38% of confident answers below the 0.8 confidence threshold that routes them to human review. Applications built on RLCD models should therefore treat the state as untrusted, argued input. Project website: https://JevAdvBench.github.io/JevAdvBench/

We empirically investigate Question Under Discussion based modelling in naturalistic dialogue by studying whether the salience of generated potential questions predicts their subsequent resolution. Building on Wu et al. (2024), we construct a dataset of 7,124 questions automatically generated from utterances and preceding context from the British National Corpus, and annotated for salience and answerability. We find a robust but low positive correlation between salience and answerability in dialogue, indicating that more salient questions are more likely to be addressed. However, this effect is markedly weaker than in monologic text, suggesting that conversational structure is less predictable. We further observe that structured interactions exhibit stronger alignment between annotators than less organised dialogues.

Activation steering is a powerful training-free paradigm for controlling large language models at inference time. However, standard approaches estimate a per-layer steering direction from contrastive data and apply it on the layer's entire representation space, which may couple the intervention to off-target properties present in the contrastive data and degrade unrelated capabilities. To mitigate this issue, we introduce LocUS (Localized Unembedding Steering), a method which grounds activation steering to the model's own output vocabulary subspace. By identifying a property-specific linear subspace within the unembedding matrix, LocUS enforces a geometric constraint that restricts the steering transformation to a specific subspace and at the same time localizes its application to a sparse subset of attention heads. Extensive evaluations across three model families on toxicity mitigation, sentiment redirection and sycophancy suppression show that LocUS matches or outperforms state-of-the-art baselines while intervening on under 6% of parameters and better preserving general capability.

Patient-facing AI assistants promise valuable support to patients, but incoming queries can pose medical risks. To create guardrails, we work with oncologists to define three ordinal risk axes: Medical Urgency, Psychological Urgency, and Topic Sensitivity. We propose Clinical Guardrail Probes (CG-Probes) to measure the risks from query embeddings. We probe for each axis in the normalized embedding space of frozen embedders via the difference-in-means method, treating each axis as a potential linear direction. To train the probes, we cluster 79,658 Czech oncology search queries with BERTopic and use these clusters to generate pairs of queries with contrastive risk levels via few-shot prompting. We evaluate the approach on 200 queries (90 real, 110 synthetic), each graded by two oncologists, against two open-weight LLMs and a frontier LLM. We find that urgency-based axes are recoverable as linear directions, and the probes are competitive with open-weight LLMs (no significant differences in quadratic-weighted kappa) at a fraction of the latency. Each axis yields a scalar score that clinicians can inspect and use to set escalation thresholds. The pipeline requires only search logs, axis definitions, and black-box access to the embedding model, suggesting transferability across healthcare domains. Robust validation on new queries and axes remains future work.

Collaborative science learning requires nuanced interpretation of student dialogue to characterize how learners identify knowledge gaps, build explanations, and work toward resolution - a theory-driven analysis that is labor-intensive and difficult to scale. We investigate whether instruction-tuned large language models (LLMs) can support multidimensional analysis of collaborative sensemaking without task-specific training, and whether structured knowledge-state information improves model inference. We evaluate two mid-size LLMs on 23 richly annotated, expert-labeled episodes across prompting conditions that vary definitional scaffolding, reasoning mode, and turn structure. Without reasoning, models tend to overpredict successful sensemaking; reasoning-enabled prompting improves identification of unsuccessful cases. Knowledge-state diagnostics provide additional grounding, improving detection of unsuccessful sensemaking and increasing agreement with expert annotations. No single configuration performs best across all sensemaking dimensions, underscoring the multidimensional nature of the task.

Conversational agents, generative recommenders, and personalized advertising all rest on one capability: understanding each user from raw behavior. Prevailing industrial practice is task-specific: for each task, a relevant subsequence is extracted from the full history and a dedicated model trained on it. In production it hits two bottlenecks. First, even after filtering, a single-task sequence stays extremely long: content-interest summarization reads several hundred items per user, tens of thousands of tokens once serialized as prompt text. Second, profiles are refreshed routinely: a billion users weekly, roughly 100K QPM in aggregate, which under a fixed GPU budget sets a hard throughput floor. Compression is therefore mandatory, yet truncation or coarse compression can silently distort the profile, introducing four hallucination types (fabrication, omission, date misattribution, broken logic) that, with no way to evaluate the compressed representation itself, surface only as diffuse degradation in downstream metrics. We present KuaFu, a unified behavior-compression layer whose minimal unit is one behavior item. A two-axis projector compresses each item into 2-4 tokens of width 128-256 (about 10x along the token axis, 20x along width; per-item cache 10 KB to 0.5 KB), with fidelity-oriented four-stage training and layered intermediate evaluation. Across four production profiling tasks it matches or exceeds uncompressed single-task production models on all five headline metrics, raises per-GPU throughput by 37%-350%, and saves 190 GPUs. On public benchmarks it nearly always beats prior compressors at the same compression ratio (up to +17.7 EM on out-of-domain MRQA); on RecBench, a 4B model surpasses its 8B counterpart by 1.90 points. KuaFu has run on the Tencent advertising and recommendation platform for ten months, lifting overall GMV by 1.37%.

Researchers increasingly use generative large language models (LLMs) to convert corporate text into empirical variables. We examine the extent to which LLM-based textual measures are invariant to model choice using thirteen measures, including sentiment, management clarity, uncertainty, answer specificity, and climate and political risk. Seven LLMs from different providers score earnings call transcripts of S&P 500 companies on these constructs. Cross-model rank correlations average only 0.52, and transcript-level differences common across providers account for only 34% of total score variation. Cross-model disagreement does not predict subsequent analyst or market disagreement, consistent with a substantial model-specific component rather than common ambiguity in the underlying disclosure. Model choice significantly affects downstream inference, with coefficient magnitudes, signs, and statistical significance varying substantially across models. Averaging across providers makes transcript rankings more stable for most constructs, but score levels remain sensitive to the models included in the ensemble. LLM-generated variables should therefore be treated as model-contingent measurements and validated across providers.

Post-training quantization (PTQ) is a practical approach to reducing the memory and computational footprint of large language models (LLMs) without retraining. GPTQ-based methods have become the de facto standard, yet they suffer from two complementary limitations. Methods with local, layer-wise objectives lack global supervision; while methods with global objectives fix their Hessian estimates at the start and ignore first-order gradients, so their guidance grows stale as quantization proceeds. This paper presents G$^2$PTQ, a unified PTQ framework with Generalized Gradient Compensation that integrates both first- and second-order information under a globally supervised, block-wise optimization objective. By refreshing gradient and Hessian estimates before quantizing each Transformer block, G$^2$PTQ avoids the staleness of prior global methods. Furthermore, to stabilize the exact first-order compensation, we introduce a trust-region scaling mechanism that dynamically bounds the gradient step to prevent exploding weight updates. Finally, we derive efficient implementations for block-wise Hessian approximation and exact gradient compensation. Experimental results on various model families and bit-widths demonstrate that G$^2$PTQ enables better alignment with the full-precision model, outperforming state-of-the-art baselines. Code is available at: https://github.com/G2PTQ/G2PTQ.

Open rerankers trained for general web retrieval transfer imperfectly to e-commerce, where ranking decisions depend not only on topical relevance but also on user preferences, product constraints, and comparative product fit. These preference signals are difficult to supervise at scale: real search traffic provides authentic queries and candidates but no clean pairwise labels. We present ZooWork-ShopRanker, a family of e-commerce rerankers (0.6B, 4B, and 8B) aligned to judge-labeled shopping preference. Training pairs are labeled by a panel of reasoning large language models (LLMs) from different families acting as a preference oracle, with position-debiased judgments and agreement tiers, and the rerankers are trained on these labels. The aligned 8B flagship then serves as a distillation teacher for the efficient 4B and 0.6B models, which are fit to its scores and sharpened on judged pairs. To measure progress, we introduce ShopRank-Bench, a contamination-limited benchmark of ~10,000 private-traffic preference pairs in both text formats, tiered by how many judge families committed to each label. ZooWork-ShopRanker-8B and -4B significantly outperform the strongest open reranker baseline, every model significantly beats its own un-aligned base, and ZooWork-ShopRanker-0.6B beats its size peer; the gains hold in both formats and extend to common MTEB benchmarks. We release the models and the dual-format ShopRank-Bench to facilitate further research.

As large language models increasingly mediate information access, factually accurate and independent answers are critical. However, these models can exhibit sycophancy by aligning their responses with users' stated beliefs even when those beliefs are incorrect, potentially presenting misinformation as independently verified and reinforcing users' confidence in false claims. Prior work leaves unresolved whether introducing user beliefs causes correct responses to become incorrect or uncertain, or causes uncertain responses to become belief-aligned incorrect answers. It also remains unclear whether anti-sycophancy interventions preserve or restore factual accuracy or merely shift responses toward uncertainty. We analyze factual sycophancy in Chinese-language information seeking using yes/no fact-checking questions. Our analysis covers 364,941 responses from three frontier Chinese-based LLMs (DeepSeek, Qwen, and Doubao) to 12,165 factual questions derived from real-world Chinese search queries. We evaluate the models with and without reasoning across baseline, belief-conditioned, and anti-sycophancy prompting, tracing matched shifts among correct, incorrect, and uncertain responses. Under incorrect user beliefs, we distinguish belief-aligned errors from losses of factual confidence, in which initially correct answers become uncertain. Patterns vary across models and reasoning settings: reasoning is not a consistent safeguard, and anti-sycophancy instructions can reduce incorrect agreement while increasing uncertainty. In Chinese-language factual question answering, avoiding agreement with false beliefs is therefore not equivalent to preserving factual accuracy, highlighting the value of transition-level evaluation. Such behavior may undermine the reliability of LLM-mediated information access by reinforcing misinformation or weakening users' confidence in factually correct answers.

Text-to-speech needs written text in spoken form, and speech recognition output needs the reverse. For Khmer, neither direction has a maintained open-source tool, and the script makes both harder: words are not separated by spaces, and number words occur inside ordinary words. We present Tha, a Khmer text normalization and inverse text normalization toolkit built from weighted finite-state transducers. It segments and classifies a whole line in one shortest-path search, and a second transducer rejects token boundaries inside a Khmer syllable. On Google's Khmer test suite, Tha agrees with the reference on all 274 cardinals up to one spelling variant, and on 2,906 real TTS prompts, 153 of the 158 sentences it rewrites are correct. Tha is open source under the Apache 2.0 license.

Uniform discrete diffusion models (UDMs) commonly use explicit time conditioning, but we find that it can often be unnecessary in practice. In this paper, we first show that the population-optimal UDM predictor generally depends on time: time controls how much the model should trust the observed context. We then show that this dependence can become negligible in finite-data settings relevant to language. When a corrupted training sequence remains much closer to its original clean sequence than to competing training sequences, the empirical-optimal predictor is nearly insensitive to time over most of the diffusion trajectory, where the guarantee weakens toward the high-noise endpoint. Empirically, trained language UDMs exhibit limited time sensitivity over most of the trajectory, while time-agnostic predictors remain competitive with, and often outperform, time-conditioned models across datasets and training objectives. These results challenge the use of explicit time conditioning in UDMs: although the population optimum depends on time, explicitly conditioning on it may often be unnecessary in practice.

Lexical semantic change is usually summarized by a scalar distance between independently sampled period distributions. This measures how much a word changed, but does not reveal when it changed, which mechanisms and component movements carried the change, or which usages support the attribution. We introduce Coupled Usage–Sense Processes (CUSP), which derives these answers from a single marginal preserving temporal process. A hierarchical coupling relates contextual distributions through latent usage components, while Markov composition makes adjacent and longer span correspondences compatible. Displacement operators quantify change magnitude and timing, split variation exactly between movement of component centers and reorganization within components, and attribute it to transported component pairs. Word-local modes resolve distinct directions of change and their activity over time, while representative passages from attributed components ground the analysis in text. Under a Gaussian mixture specialization, we prove parametric recovery of the operators and squared distances. Synthetic experiments support the predicted rate. CUSP remains competitive on English and German DWUG and recovers controlled Janus profiles while maintaining compositionally coherent transport. A large corpus of US court opinions demonstrates transition, mode, and passage attribution in unlabeled natural text. CUSP thus makes magnitude, timing, mechanism, movement, modes, and textual evidence compatible views of one lexical history.

Large language models are increasingly used as personalized writing assistants, but adapting a model across many authors can compromise individual writing style by pulling author-specific signals toward a shared register. Federated parameter-efficient fine-tuning (PEFT) offers a data-local setting for this multi-author adaptation problem: clients keep author text local while sharing compact adapter updates. However, we show that standard aggregation can preserve continuation utility while making different authors' generations less distinguishable in style space, a failure mode we define as author-style homogenization. We evaluate author-style retention with Angular Style Classification Encoder (ASCE)-based diagnostics on our main BlogText benchmark and ASCE-independent external authorship verification. Using this protocol, we find that common federated PEFT baselines can preserve semantic utility while averaging out author-specific signals. To address this homogenization, we instantiate FAVoR (Federated Authorial Voice Retention), an author-style residual mechanism for federated PEFT. FAVoR uses a shared-private adapter design: clients upload shared-adapter updates while retaining author-specific residual corrections locally. Across BlogText and external Mythos-Reddit validation, FAVoR improves author-style retention over standard and personalized federated PEFT baselines. These gains come with small continuation-utility trade-offs and are supported by component ablations, external verification, and cold-start transfer.

Continual fine-tuning is essential for large language models (LLMs) to dynamically adapt to real-world environments, yet it inevitably suffers from catastrophic forgetting, particularly the performance degradation of previous tasks and LLMs' general-purpose knowledge. Although existing methods, such as orthogonal gradient projection, mitigate the forgetting across various fine-tuning tasks, they fundamentally fail to preserve pre-training LLMs' inherent general-purpose knowledge because the original data and gradients of off-the-shelf pre-training LLMs required by these methods are strictly unknown and highly diverse. To bridge this critical gap, we propose EoupCT, a novel framework designed to Estimate and Orthogonalize Unknown Pre-training gradients for Continual LLM fine-Tuning. Specifically, EoupCT estimates pre-training gradients by dynamically generating pseudo data that is most susceptible to forgetting for new tasks through a learnable soft prompt equipped with Gumbel-Softmax relaxation. Furthermore, we formulate a multi-objective optimization problem and introduce a first-order efficient Pareto optimizer that jointly optimizes LLM parameters and the soft prompt, rigorously enforcing orthogonality between new task updates and the estimated pre-training gradients. Extensive experiments across multiple LLMs demonstrate that EoupCT effectively preserves both task-specific proficiency and inherent general-purpose knowledge, successfully mitigating the catastrophic forgetting.

Accurate and efficient pronunciation transcription is essential for preparing text-to-speech training data at scale. Existing approaches have different limitations: grapheme-to-pronunciation (G2P) and speech-to-pronunciation (S2P) methods each capture only partial information, using only text or only speech, while speech-and-text-to-pronunciation (ST2P) methods use both but require costly pronunciation-annotated data. To address this problem, we propose a training-free ST2P pipeline that integrates both lexical and acoustic information at inference time. Lexical resources and G2P tools generate text-constrained candidates, and a left-to-right greedy search selects the best one using whole-sequence negative log-likelihoods from frozen pretrained S2P models. On three Japanese corpora, our method reduces Character Error Rate (CER) from 0.60–1.40% (text-only baseline) to 0.04–0.17% with reference transcripts, and 0.64–1.58% with ASR transcripts. It outperforms all baselines, including a trained ST2P model and commercial multimodal LLMs. Our greedy search method is 3–3.5$\times$ faster than beam search at similar CER, and the cascade is 2$\times$ faster than direct decoding ensuring the efficiency and accuracy. In Spanish, French, and preliminary English, it also surpasses four open multimodal LLMs and the best traditional methods.

Quantum-inspired algorithms emulate quantum mechanical principles, such as, superposition, interference, and probabilistic amplitude evolution, on classical hardware by representing candidate solutions as qubit vectors and evolving them through rotation-gate operators. This approach offers higher optimization performance without physical qubits, and has been shown to achieve order-of-magnitude speedups (10–80$\times$) over traditional solvers on combinatorial, high-dimensional NP-hard problems. A critical barrier to adoption, however, is the lack of a unified execution framework that delivers both algorithmic performance and hardware portability. We present Cross-Backend Quantum Inspired Evolutionary Optimizer (QIEO), the runtime core of BQP's BQPhy solver, which addresses this gap through a single-source-of-truth architecture. One C++ implementation of the QIEO algorithm is compiled once per hardware target and exposed to multiple high-level languages via thin binding layers. The framework dispatches to CPU (sequential), OpenMP~5 (multi-core), CUDA (NVIDIA), and HIP (AMD) backends at runtime, adapting kernels to each device's memory hierarchy and warp/wavefront execution model. The framework's real-world utility is validated through binding demonstrations that share the identical C++ runtime. BQPhy's Python library is demonstrated on a neural network hyperparameter optimisation achieving 88.60% test accuracy on MNIST. BQPhy's MATLAB's Toolkit is tested on wind farm layout optimisation attaining $365\,399 \pm 4\,552$~MWh/yr, which is statistically indistinguishable from particle swarm optimisation and $+7.6\%$ above genetic algorithms on a 32-variable constrained engineering problem. The Julia package tackles the Lotka–Volterra parameter estimation where BQPhy replaces native Julia solvers on the same residual, cutting mean SSE by $2.1\times$.

Large language models (LLMs) excel at natural language processing but struggle to interact with external environments. Tool learning provides a promising way to extend LLMs into actionable agents, where tool selection is a critical prerequisite for successful tool use. Existing work often assumes a small or predefined set of tools, leaving large-scale tool selection underexplored. Real-world repositories contain a vast and diverse array of tools, making it difficult for LLMs to effectively search, distinguish, and compose tools under context-length constraints. We identify large-scale tool selection as a new challenge for agentic reinforcement learning, highlighting that existing RL methods for knowledge-based question answering are inadequate for selecting tools while considering compatibility. To address this challenge, we propose ToolSearcher, a novel RL framework for effective multi-turn search and fine-grained optimization in large-scale tool selection. Specifically, we introduce category-constrained tool discrimination to improve the model's ability to distinguish functionally similar tools, event-level search modeling to explicitly optimize the discovery of target tools during multi-turn search, and trajectory-aligned credit allocation to provide fine-grained reward signals for different stages of the search-selection process. Extensive experiments on large-scale tool selection benchmarks demonstrate that ToolSearcher consistently outperforms a set of strong baselines in challenging settings involving iterative search and complex tool composition.

How should we evaluate language models when more than one interpretation can be right? Cultural benchmarks often test factual knowledge, agreement with survey responses, or recognition of a predefined meaning. These tasks leave open whether a model can explain how a cultural reference works in a particular text, support a reading with evidence, or revise it after criticism. This is a question of interpretive depth, complementary to the breadth of cultural coverage. We argue that literary interpretation offers a useful setting for studying these capabilities. We focus on cultural referencing and reuse: how texts invoke, repeat, and transform earlier expressions across historical and linguistic contexts. Our central claim is that literary scholars can disagree about an interpretation while recognizing the quality of its support. We propose linking evidence-centered benchmarks, evaluation that preserves scholarly disagreement, and model-development experiments on literary data, contextual resources, and scholarly feedback. Danish literature provides a concrete starting point, with implications for other languages and domains. The aim is to develop alternative evaluation strategies that go beyond conventional benchmark metrics and guide model development toward cultural robustness in AI systems.

Large language models (LLMs) are increasingly used to assess long-form medical videos, but their effectiveness may depend on whether transcripts are provided in full or compressed through summarization, retrieval, or claim screening. This study examines how such transcript compression affects LLM-based veracity classification of Japanese medical YouTube videos. We compare four transcript input designs: full transcripts, LLM-generated summaries, RAPTOR-based retrievalaugmented generation (RAG), and Screening, which extracts candidate medical and health-related sentences. Using 74 long-form videos labeled as Real or Fake, we evaluate classification performance and analyze linguistic changes using J-LIWC, hedge expressions, and institutional or technical terms. The full-transcript Baseline achieved the best performance, whereas all compressed inputs increased false negatives, meaning that Fake videos were more likely to be misclassified as Real. Summary caused the largest performance drop, while Screening performed best among the compressed inputs but still omitted many medically relevant sentences. Linguistic analyses showed that these errors were not explained by a simple increase in certainty. Instead, Summary reduced affective, social, temporal, cognitive, and conversational cues, while Summary and RAG made institutional and technical terms more salient. These findings suggest that transcript compression can represent Fake videos as more coherent and authoritative inputs, thereby weakening cues needed for misinformation detection

Difference-in-differences (DID) studies are widely used to evaluate climate policy, but assessing the evidence supporting their identification assumptions remains challenging. We introduce ARGUS, a structured language-model pipeline that audits reported evidence against an eleven-dimension assumption-implication-evidence rubric and abstains when relevant evidence cannot be retrieved. We evaluate ARGUS using injected flaws, economics papers, and a small pilot with reconciled labels. On the 11-flaw benchmark, ARGUS detects 73% of planted flaws, compared with 18% for a keyword-based pipeline. Across 26 economics papers, ARGUS abstains on about 40% of paper-dimension assessments for lack of retrievable evidence. In a five-paper pilot with labels reconciled by two annotators, it assigns a higher risk level than the labels on 25 of the 33 assessments it completes. A rule fixed before the labels arrived removes most of this in-sample; weighted agreement stays low. ARGUS provides evidence-linked risk reports that localize potential weaknesses for expert review, without adjudicating causal claims. Code and data: https://github.com/yonghongzhang-io/ARGUS

Black-box On-Policy Distillation (OPD) seeks to improve a student from its own generations when the teacher provides sampled responses but not token probabilities. Adversarial distillation offers one route: it learns a discriminator over prompt-matched teacher and student responses and uses its score as the policy reward. However, sampling discriminator negatives from the latest student at each step couples the learned reward to a negative distribution that changes after every policy update. We address this moving-target problem with persistent-negative adversarial distillation, a live-pool method that replaces a fraction of each discriminator batch with historical, prompt-matched teacher–student comparisons. Under matched discriminator compute, historical comparisons train the discriminator, while GRPO remains on-policy with fresh student responses. Our analysis identifies the Bayes-optimal reward as a teacher-to-negative log-density ratio and, under explicit assumptions, shows how persistent negatives anchor the discriminator and reduce reward-estimation MSE relative to fresh-negative training. Across two student families, three judges, and four judged-chat benchmarks, persistent-negative adversarial distillation consistently improves performance over current methods at matched discriminator compute. It also yields smoother fresh-policy discriminator trajectories, with fewer below-chance dips. These findings identify the discriminator's negative distribution as an important design axis in black-box on-policy distillation.

Prior work has examined the self-consistency of LLM-generated explanations using surface-level perturbation methods. However, the strength of these perturbations is not explicitly measured and controlled. In this work, we propose an LLM-as-a-judge approach to measure perturbation strength in a unified manner across input and CoT perturbations. We then evaluate the self-consistency in explanations generated from various LLMs under controlled strength conditions, ensuring a fair comparison across perturbation types. Experiments show that our proposed LLM-based perturbation strength measure outperforms other embedding- and probability-based approaches and that input perturbations generally affect LLMs more strongly than CoT perturbations. Our work suggests that judgments about a model's self-consistency is fair only within the same perturbation type.

In this paper, we propose I-Parakeet, an integer-only implementation of NVIDIA's Parakeet-CTC (0.6B parameters) that runs on a smartphone NPU without any floating-point operator or CPU fallback. Modern Conformer ASR models are hard to deploy on edge devices because of their size, and quantized models still fall back to floating point for numerically sensitive operations. This prevents them from fully exploiting integer accelerators such as mobile NPUs. To achieve this, our contributions are threefold. First, we derive an integer formulation of the relative-positional self-attention at the core of the Conformer. We fuse its two score branches with different quantization scales and the relative shift into integer-only operations. Second, we introduce a minimax-optimized Swish approximation that minimizes the maximum error of the Swish output. Third, a layer-wise range analysis of activations yields two targeted remedies: an INT16 grid for the BatchNorm output and percentile calibration for the heavy-tailed pre-encoder activations. I-Parakeet achieves 4.97% WER on LibriSpeech test-other, running on a Qualcomm NPU at a real-time factor of 0.048, 7.5x faster than a CPU baseline.

Looped transformers reuse weights across recurrence steps, making low-bit quantization especially attractive. We identify two distinct failure modes of standard post-training quantization. On Huginn-3.5B, per-channel INT4 fails primarily at the non-residual loop-entry adapter, while quantizing the residual core is much less damaging. We call this feedback exposure: a quantized layer perturbs the recurrent state without an identity path, and the resulting error is fed back at later steps. Controlled experiments on linear filters and Mamba state-space models show that feedback exposure also occurs outside transformers. Grouped INT4 reveals a separate failure, calibration blindness: our one-step GPTQ baseline builds its Hessian from step-0 activations, leaving input directions used later in the recurrence nearly unweighted. Across nine checkpoints from seven looped architectures, one-step GPTQ is worse than round-to-nearest (RTN) on the primary task metric for five checkpoints. Accumulating the GPTQ Hessian across recurrence steps outperforms both one-step GPTQ and RTN on all nine checkpoints and recovers bf16-level accuracy on Huginn. These results separate two questions for PTQ on looped models: where quantization error enters the recurrence, and which states calibration sees.

Prior research has demonstrated that the choice of prompt template during Supervised Fine-Tuning (SFT) significantly impacts the robustness of safety alignment afterwards. However, the influence of template selection during Knowledge Distillation (KD) from teacher to student remains largely unexplored. Thus, we fill this gap by analyzing how different template configurations influence the pre-existing safety alignment of the student. We observe a significant degradation of safety alignment present in the aligned base instruct-tuned model. Specifically, we find that utilizing chat templates renders the model more compliant with harmful queries compared to a non-chat template. These findings are consistent across three models: LLaMA, Gemma and Qwen model families and are evaluated across multiple safety benchmarks. We further show that using a non-chat template during distillation better preserves the base student's internal representations, while chat template distillation induces a larger representational shift. Code: https://github.com/anjilab/role-of-prompt-template-in-kd

This paper proposes an architecture for equipping large language models (LLMs) with audio-understanding capabilities without fine-tuning their weights. The proposed symbiotic architecture employs an injector module that writes audio-conditioned vectors directly into the target LLM's short-term memory, i.e., the key-value (KV) cache, enabling the LLM to behave as an audio language model (ALM). The architectural advantages are twofold. First, it improves the scalability of ALMs: because the proposed method bypasses the LLM during audio injection, the injection cost is governed by the injector width rather than the backbone width, and can therefore scale more slowly than the cost of full-backbone prefilling. Second, since the training scheme does not update the LLM weights, the original capabilities of the LLM are preserved without the risk of degradation from fine-tuning. The effectiveness of the proposed method is evaluated on both audio-understanding tasks (automatic speech recognition, audio question answering, and acoustic scene classification) and text-only tasks. We confirm that, while activating fewer parameters during audio prefilling, our architecture outperforms the conventional method with a frozen LLM and approaches the performance of a fine-tuned ALM, all while preserving the backbone LLM's original text-only task performance by construction.

Tandem speech-to-speech architectures couple a responsive speech frontend with an asynchronous text backend. In KAME, a large language model (LLM) serves as the backend, supplying candidate responses as guidance to the speech frontend while the user is still speaking. Ordinary conversation recordings capture the eventual response but not the guidance the backend would supply during the user's utterance. Generating the missing guidance with a simulator LLM adds substantial data-preparation overhead when training on real conversations. We propose randomized intermediate guidance, which derives guidance directly from the conversation corpus rather than simulating backend LLM behavior. During training, target responses provide informative guidance, while randomly sampled responses provide potentially irrelevant updates during the utterance. This combination aims to teach the frontend to use backend information selectively. On synthetic dialogues, KAME trained with this recipe achieves response quality comparable to that of the LLM-generated and similarity-based baselines. Training on 3.8k hours of real conversations improves smooth turn-taking and audio-judge naturalness over synthetic-data KAME while retaining a response-quality advantage over Moshi. These results show that randomized guidance offers a practical route to combining the response-quality benefits of tandem models with natural conversational behavior learned from real speech.