Papers

Multilingual text-vision embedding models are essential for cross-lingual image-text retrieval, but Southeast Asian languages remain poorly supported due to the region's linguistic diversity and limited data and computing resources. In this paper, we introduce SEA-CLIP-Tiny, a compact multilingual text-vision embedding model for Southeast Asia with fewer than 50M parameters. Our model adapts a CLIP-KD-style framework to Southeast Asian multilingual settings through regional data curation and multilingual teacher guidance. Experiments across seven Southeast Asian languages show that SEA-CLIP-Tiny achieves the strongest average retrieval performance among the evaluated student models, reaching 12.9%, 31.5%, and 42.2% at R@1, R@5, and R@10, respectively. Compared with MobileCLIP2, it improves average R@10 by 12.1 points while using 38.4% fewer parameters and lower measured CPU latency. These results highlight the importance of region-aware training for efficient multilingual text-vision models in Southeast Asia.

KV cache eviction methods such as SnapKV and PyramidKV rank tokens solely by mean attention over a small observation window. We study a unified score, $\mu_i+\lambda_1\sigma_i+\lambda_2\mathrm{corr}(i,S)$, adding attention dispersion across window queries and redundancy relative to selected tokens. For $\lambda_2<0$, the score penalizes similarity to selected tokens as in maximal marginal relevance (MMR), without extra forward passes. To test whether this relevance-diversity balance should vary with depth, we compare fixed global coefficients with three-segment and quadratic profiles. Only these depth profiles are searched on a development split under a $\sinh$ reparameterization. On all 16 English LongBench datasets with Mistral-7B at a budget of 64 entries per layer, a single global diversification constant improves 13 of 16 datasets (macro +1.1); the gain holds at budget 32 and narrows at 128. Per-dataset search finds no detectable layer structure on most datasets; on passage retrieval it finds a large one: a mid-layer sign flip that rewards similarity and is worth +9.6 over the baseline at budget 64 and, without re-tuning, +13.2 over the global constant at budget 128. Ablations attribute the gain to the redundancy term; replaying every accepted search state on the held-out test set separates genuine structure from tuning noise.

When a language model receives two conflicting documents as input, how does it decide which one to prioritize? Does it rely on how the sources are framed or the presentation order of the documents? We evaluated this behavior on Google's pre-trained Gemma 4-e4b model across a targeted behavioral suite (n = 13 items, 784 forward passes in short, single-turn contexts) using a completely counterbalanced experimental design. This setup allowed us to mathematically isolate the specific effects of source framing and reading position, while ensuring the model's natural vocabulary biases were canceled out. Across ten test conditions, we discovered the following: 1. Source framing heavily overpowers reading position. When directly competing, the semantic framing of a source (such as presenting it as an official guideline or a fresh update) had a significantly stronger impact on the model's final answer than the presentation order of the document. 2. The model favors the first document it reads, but this bias is highly variable. While the model consistently demonstrated a primacy effect (preferring the first document presented), the actual strength of this bias fluctuated by at least a factor of 5 based solely on the surface wording. 3. Overall structural repetition, not short copy-cues, drives positional bias. The model's preference for the first document is not a mechanical reaction to short, repetitive trigger phrases, such as "is [Answer]". However, the primacy effect does increase significantly when the two competing documents are structurally identical, using word-for-word verbatim templates. Introducing variation in the overall wording between the two sources reduces this positional bias.

"System One" decision models such as TypeSafe's Jev and its open counterpart Laya answer typed questions about a text in a single forward pass with calibrated probabilities, but they cannot ask for missing information: when a first message does not say what separates two departments, they guess. We present LAVOIR (Laya with Value-Of-Information Routing), which places the candidate pieces of missing information (slots) in the input next to the answer options, so that one forward pass returns both the decision distribution and, for every slot, the expected gain in the probability of the correct decision if the user were asked about it. VOI targets need no human labels: gold decisions come from schema rules, an LLM only verbalizes messages and answers, a model from another family checks every text, and pairing each message with several profiles makes regression on realized gains estimate the expected gain. A Gini-impurity cap bounds the predicted value by what a calibrated model can still gain. In a controlled study, decisions on seen schemas are statistically indistinguishable from the Bayes ceiling. The final model's question policy matches a greedy oracle VOI policy on seen schemas (AUC 0.799 vs. 0.797), and with at most 0.5 questions per conversation it is 14.1 points more accurate than never asking. On real ABCD conversations, one real exchange raises accuracy by 8.3 points where LAVOIR asks and leaves it unchanged where it does not; on SGD the cap lowers the asking rate from 93% to 8.6%. On Laya's twelve benchmarks LAVOIR is above Laya's reported scores on seven, and it answers a question in 31 ms (median, GH200).

Streaming video understanding requires models to interpret evidence as it arrives, yet current evaluations often report task scores without specifying when evidence becomes valid, how visual history is maintained, or how responses are triggered. As a result, similar scores may correspond to different workloads, failure modes, and operational behavior. We introduce TRACE (Temporal Audit and Condition-aware Evaluation), a condition-aware benchmark and evaluation framework that makes these factors explicit. TRACE combines temporally audited visual tasks with evidence timing and instruction-dependent trigger annotations, a unified causal Core–Adapter protocol that controls information availability while recording actual history processing and response events, and multidimensional reporting of answer quality, timeliness, response-selection behavior, workload, completion, and reliability. On 1,240 records from 517 videos, we evaluate eight publicly available models or systems in eight configurations. We find that nearly identical QA accuracy can mask substantial differences in completion, answer validity, and generation workload, while proactive performance separates into response quality, response delay, false alarms (responses emitted while no target window is currently valid and a later one remains), and missed target windows. These results show that streaming-video performance should be interpreted as execution-conditioned system behavior rather than a single score. Our benchmark and code can be accessed at \href{https://github.com/om-ai-lab/trace-bench}{https://github.com/om-ai-lab/trace-bench}.

Large language model email assistants are particularly vulnerable to indirect prompt injection because untrusted email content can be retrieved into the model context and influence subsequent tool use. Existing prompt injection detectors mainly formulate this problem as binary malicious text classification, which overlooks the important factor that harmful agent behavior often arises through a sequence of stages. We propose a detection framework that models this attack chain by combining a text detector, verifiers specific to each stage, explicit rule-based risk signals, user intent and action consistency analysis, and a logistic decision policy. To support this framework, we derive attack chain labels from prompt injection datasets, evaluate the proposed framework under random splits, temporal phase transfer, conditional stage transfer, cross-dataset transfer, and conduct ablation studies on multiple benchmarks. Results show that random train test splits substantially overestimate robustness under distribution shift, while later tool argument stages are more predictable than earlier stages in the framework. We also show that training on harmless emails that resemble attacks helps reduce false alarms while preserving the ability to detect real attacks. Across five binary benchmarks, our framework achieves a mean F1 score of 0.406 under the strict threshold setting policy, compared with 0.216 for the strongest of five pretrained detectors evaluated without additional training. These results highlight the value of combining attack stage predictions with checks for conflicts between the user's request and instructions in retrieved emails. Our experiments also demonstrate the importance of training with challenging benign examples to balance attack detection and false alarms.

On-policy distillation (OPD) trains a student model by having it generate trajectories, then matching its next-token predictions with an external teacher's next-token predictions. This provides dense, token-level supervision to the student. On-policy self-distillation (OPSD) eliminates the need for the external teacher. Specifically, a second frozen copy of the student model, now given the ground truth in its context, serves as the teacher. The student model only receives the problem and learns to mimic the privileged teacher model, while the teacher remains frozen throughout training. Previous work showed that freezing the teacher is useful for training stability, but we argue that this can prevent the teacher from incorporating the improvements learned by the student during training. Our primary contribution is to address this limitation with a recursive framework built around two complementary components. First, we let the privileged teacher co-evolve with the student so that revision learned in one round can guide the next, a process we refer to as Dynamic Co-Evolution (DCE). Second, because stronger revision can also make responses too verbose and self-critical, we additionally train on shorter, verified rewrites of the model's own on-policy responses. We call this complementary objective Self-Refined Concise Learning (SRCL). Overall, our comprehensive evaluations show that DCE+SRCL outperforms OPSD across multiple model scales and four competition-level mathematics benchmarks. Specifically, on Qwen3-8B, DCE+SRCL reaches 65.97% Average@12, outperforming OPSD by 35.62 percentage points while reducing mean output length by 7.80% relative to DCE alone.

Personalizing text-to-image diffusion models from a few reference images requires preserving subject identity while following prompts that describe new contexts. Full-model fine-tuning is parameter-intensive, whereas low-rank adaptation (LoRA) reduces the number of trainable parameters but leaves open how adaptation capacity should be distributed across layers. We introduce Diffu-LoRA, a parameter-efficient method that learns this allocation through gated low-rank adaptation. Diffu-LoRA inserts trainable low-rank components into the linear layers of Transformer blocks and assigns a learnable gate to each component. Bilevel optimization updates the adaptation weights and gate parameters on separate data splits, while progressive pruning removes components with the lowest gate values to meet a prescribed rank budget. This procedure allocates adaptation capacity nonuniformly across layers while keeping the pretrained backbone frozen. Experiments with Stable Diffusion on subjects from DreamBooth and additional collected datasets show improved overall subject fidelity and prompt alignment relative to the evaluated fine-tuning baselines. Ablation studies examine the contributions of bilevel optimization, progressive pruning, and adapter placement. These results support learned rank allocation as a practical approach to parameter-efficient diffusion model personalization.

Spatio-temporal forecasting is a cornerstone of logistics, urban planning, and intelligent transportation systems. However, constrained by deployment costs and maintenance resources, sensor networks often lack comprehensive spatial coverage, rendering Forecast Unobserved Node States (FUNS) a critical yet formidable challenge. Conventional models rely on historical observations and typically falter when encountering nodes without prior records. To address this, we redefine the problem as a conditional generation task on spatio-temporal graphs and propose GenST, a framework that introduces Large Language Models (LLMs) as a semantic bridge, leveraging a pre-trained LLM fine-tuned to extract rich semantic features from node descriptions, such as functional zones and road network structures, to compensate for missing spatio-temporal signals. Specifically, we design a two-stage generative architecture: a Spatio-Temporal VAE first compresses spatio-temporal dynamics into a latent space, followed by a Generative Transformer (GenT) that reconstructs the future states of unobserved nodes from noise, guided by multi-modal conditions including semantics, geographic coordinates, and neighborhood contexts. Experiments on six traffic and two non-traffic datasets show GenST significantly outperforms existing baselines in zero-shot prediction tasks, demonstrating the practical potential of semantic-guided generation for mitigating spatio-temporal data sparsity.

Selective intent routing allows an assistant to act on reliable predictions while deferring uncertain requests. Standard confidence scores primarily reflect the base model's representation, leaving an opportunity to incorporate complementary evidence without changing its decisions. We introduce a signed lexical gate that combines a sentence classifier's logit margin with a sparse lexical model's support for the classifier's predicted intent. By assigning positive evidence to lexical agreement and negative evidence to a lexically favored competing intent, the gate retains more information than either unsigned lexical confidence or a hard agreement rule. An independent binomial calibration stage selects an operating threshold for a specified risk target. Across ten runs on BANKING77, CLINC150, and HWU64, the proposed score reduces area under the risk-coverage curve by 15.8%, 15.1%, and 11.8% relative to a learned semantic-only gate. At a nominal 5% error target, it increases accepted coverage by 1.83 and 5.14 percentage points on BANKING77 and HWU64, while CLINC150 is already near full coverage. At a stricter 2% target, the simultaneous binomial procedure yields a nonempty policy in all 30 dataset-run combinations at the available calibration budgets. Matched controls show that the proposed feature improves average error ranking over the tested unsigned lexical-confidence feature, with dataset-dependent gains over binary agreement. The resulting two-feature gate provides a compact, interpretable confidence enhancement for risk-calibrated intent routing while preserving the base classifier's predictions.

Running a language model on edge hardware provides private and low-latency reasoning without a network connection, and yet the small models that fit on such devices are unreliable on the tasks computers are expected to handle well, such as arithmetic, algebra, and formal logic problems. We argue that much of this unreliability is avoidable. Many queries appearing to demand reasoning are in fact structurally deterministic and permit fast and exact symbolic solutions. Therefore, forcing a probabilistic model to approximate them sacrifices accuracy and energy for little benefit. We present a neurosymbolic router that classifies each incoming query and dispatches it to the cheapest correct solver, sending structured tasks to deterministic engines and reserving the small language model (SLM) for open-ended word problems. Instead of hand-coding the routing logic, we learn a deterministic finite automaton (DFA) with the L* grammatical inference algorithm, using the SLM as a membership oracle and labeled data as an equivalence oracle. On a Raspberry Pi 4B (8 GB RAM, no GPU), evaluated on 100 untested prompts from DeepMind Mathematics, GSM8K, and RuleTaker, learned routing attains 100% routing accuracy and 98.3% overall accuracy with a 512-token reasoning budget (93.3% on word problems), compared with 72.0% for the strongest agent baseline, Program-of-Thought, and 58.7% for a tool-calling agent given the same solvers. Since formatted queries never reach the model, the router answers them in 1-11 ms and, in its 30-token configuration, runs 8.8x faster and 2.8x more energy-efficient than Program-of-Thought.

Optimal Transport (OT) provides a principled framework for comparing and transforming probability distributions while preserving geometric structure. Recently, OT has gained significant attention in machine learning due to its ability to measure discrepancies between distributions, even when their supports do not overlap, making it effective for tasks such as generative modeling, domain adaptation, and transfer learning. Despite its success in fields such as computer vision and natural language processing, OT remains relatively underexplored in speech research. Speech signals present unique challenges, including temporal dynamics, speaker variability, noise, reverberation, and heterogeneous multimodal representations involving audio, text, and visual information. These factors often lead to distribution mismatches, where OT offers a natural framework for alignment and interpretation. This work aims to promote broader adoption of OT in speech processing by: (1) reviewing OT foundations through intuitive physical interpretations and highlighting connections to modern generative models; (2) presenting computational algorithms suitable for deep learning frameworks; and (3) demonstrating OT applications in cross-domain and cross-modal speech tasks, including speech enhancement, automatic speech recognition, language and speaker recognition, and audio spoof detection. We highlight OT's strong potential for addressing distributional variations in real-world speech applications.

General-purpose multimodal agents can write robot-control programs, but repeated exploration and model-mediated action selection can make execution slow. We study how external body knowledge, successful experience, and executable skills improve an XLeRobot controlled by GPT-6-Astra in a simulated and a physical elevator-button task. In 30 fixed-start simulation trials, complete robot geometry and camera information reduce mean completion time by 57.4% relative to a baseline with only the common control interface and no prior experience; images with synchronized action and state records reduce it by 68.6% without additional body assets. In nine paired comparisons (18 trials) at starts displaced by 10-100 cm, experience recorded at the original start reduces mean time by 58-63% relative to no experience, demonstrating generalization to the tested new starting positions. During experience experiments, GPT-6-Astra spontaneously generates a short visual-feedback program. Researcher-refactored versions reduce mean local-task time by 29-31% in 27 simulation trials. Finally, 12 real-robot trials using operator-confirmed button contact demonstrate sim2real reuse: at a shared nominal start, simulation XML assets and simulation experience reduce mean time by 53.0% and 49.9%, respectively; real experience also transfers to two new starts. These results suggest a practical way to build general-purpose manipulation experiments around GPT-6-Astra: supply machine-readable body descriptions and synchronized demonstrations, and turn useful agent-generated feedback routines into reusable skills, while the agent adapts actions from current images. We release all task prompts, trial-level experimental data, and acquired skill implementations at https://github.com/hesd10/astra-robot-sim2real.

On Jan. 30, 2026, the U.S. Department of Justice released a mixed-media collection concerning Jeffrey Epstein, including about three million pages of PDFs. We describe the Epstein Files Engine, an A.I. agent The New York Times deployed to investigate the files. The Engine translated reporter questions into Google BigQuery SQL queries across three corpora: Epstein-related releases, the Times's archive and external, Epstein-related news headlines. It used an LLM to plan queries and returned citation-rich answers a reporter could verify and trust. More than 100 journalists used the Engine, and it contributed to at least 20 published stories. We report how reporters queried it and describe Diff, our text-and-visual duplicate matching method that amplified novelty signals and allowed the Engine to surface genuinely new information. We argue that newsroom agents serve newsrooms best not as autonomous writers, but as interfaces to source material and institutional knowledge.

Existing computer-use agent benchmarks do not fully evaluate agents acting as assistants. A useful assistant retrieves information across complex, multi-step workflows, synthesizes it into artifacts (documents, presentations, spreadsheets), and navigates program interfaces to produce a coherent final product. Such workflows demand reasoning and synthesis, decomposition of complex tasks, as well as visual and spatial understanding. To study agents on workflows like these, we introduce KNOWS, a benchmark of open-ended, complex, browser-based tasks that jointly evaluate these capabilities, with each task culminating in a produced artifact. To write tasks, we develop a task design rubric and a protocol for ensuring that tasks meet the requirements. Each task is paired with an evaluator, a program that combines deterministic checks with LLM judgments to balance the richness, reliability, and automation tradeoff inherent to agent evaluation. We evaluate and analyze frontier computer-use agents and browser-based harnesses. They achieve moderate scores on partial-success metrics, but the best performer fully succeeds in fewer than 3% of our complex, long-horizon tasks. Failures on visual steps render the resulting artifacts unusable, even when agents complete more than 50% of other evaluation steps. Our results expose limitations of current agents acting as end-to-end assistants, and call for progress on tool use, visual understanding, and long-horizon reasoning.

Platform policies are increasingly tested on artificial users, making agent fidelity important. Yet convincing fake profiles could also manipulate perceived public opinion before elections. Validation has concentrated on agreement with human behaviour and has paid little attention to whether an agent behaves in line with the profile it was given. The present study profiled eight Serbian participants through a questionnaire, a deep interview, and a written self-presentation, recorded their reactions to sixty-eight social media posts, and asked four language models to predict those reactions under five prompt conditions varying profile content and instruction style. Attitudinal content improved prediction over demographic backstories by a wide margin. Agents matched their stated profiles more closely than participants matched their own survey answers, and consistency proved unrelated to fidelity once profile information was present. Instructing models to respond intuitively and immediately rather than analytically gave the highest fidelity of any condition and cut the compression of individual differences from seven times the human level to three. The advantage held on posts about topics the questionnaire never raised, where that condition reached the highest fidelity of any setup and beat a crowd baseline by a wide margin, which suggests that agents prompted this way could serve as general-purpose simulated users rather than specialists on the topics they were profiled for. Results may bear implications for the development of language models, because intuition-based setups appear better suited to some tasks than reasoning-based ones.

Agent memory systems are increasingly used to maintain long-term user preferences, task states and evolving facts, but current evaluations often collapse memory behavior into final-answer accuracy. We introduce MemProbe, a cognitive-science-inspired framework for diagnosing stability-plasticity tradeoffs in agent memory. The framework is motivated by a core insight from cognitive memory research: memory is reconstructive and shaped by interference, source reliability, reinforcement, and reactivation. MemProbe turns this insight into four reusable experimental paradigms (interference, misinformation, consolidation strength, and reconsolidation window) that manipulate when a memory should be updated, preserved, or treated as uncertain. It further decomposes correctness into behavioral profiles that reveal how systems update, preserve, attribute, and temporally organize information. We instantiate these paradigms in a 56-episode diagnostic suite and evaluate six incremental memory systems under a unified protocol. Results show that systems with similar aggregate scores exhibit distinct behavioral profiles. MemProbe provides such a diagnostic lens, turning aggregate performance into interpretable profiles of memory maintenance over time. Code is available at https://github.com/jq-ding/MemProbe.

Audience sizing is a critical component of digital marketing. It enables precise resource allocation, campaign planning, and performance optimization. Traditional approaches using skeleton audiences, sampling, or predictive modeling suffer from significant delays, estimation errors, and poor scalability over high-dimensional profile data. We present REALMS (Real-time Exact Audience sizing via LLM-based Multi-attribute Search), a conversational system for exact audience sizing deployed in production on an enterprise customer data platform. REALMS enables marketers to query massive profile stores with millions of profiles and thousands of attributes using natural language and receive precise counts in seconds. The system introduces three key components: (1) a categorical attribute retrieval mechanism using embedding-based vector search to dynamically identify relevant schema attributes without manual configuration; (2) an LLM-powered NL2SQL pipeline with template-based in-context learning for accurate query generation over complex nested schemas; and (3) schema standardization enabling industry-agnostic deployment across diverse enterprise environments. Evaluation on real enterprise data demonstrates strong recall for attribute retrieval, high SQL execution accuracy, and low latency, which enables real-time interactive audience insights where prior methods required hours.

Text-to-music systems are assessed on audio quality and on how faithfully the music follows its prompt, and the CLAP score, the cosine similarity between a music-text model's audio and text embeddings, is the standard objective metric of faithfulness. We ask how accurately that score reflects the text: when an attribute is linked to an instrument (e.g., distorted guitar), does the text embedding capture that binding? To find out, we introduce an attribute swap perturbation: the caption of a real recording is edited by exchanging exactly one property, timbre, lead versus accompaniment, or order of first appearance, between two instruments. We then test four contrastive music-text models and one large audio-language model on whether the audio scores higher against the original caption than against the perturbed one. No contrastive model distinguishes the two captions reliably. The audio-language model does better, but further experiments show that its advantage rests largely on audio-agnostic language priors. Our results thus provide compelling evidence that the CLAP score and related metrics do not capture fine-grained musical meaning or attribute bindings; their representation is closer to a bag-of-words that leaves them insensitive to meaning-changing perturbations of the caption.

Children in multilingual communities often code-switch, using multiple languages in a single utterance. Can we induce cross-lingual alignment in language models by training on code-switched text? We pretrain small decoder-only transformers on two 100M-word multilingual corpora: a base corpus formed by mixing the English, Dutch, and Chinese BabyBabelLM datasets, and a corpus generated from it by inserting word- and sentence-level code-switching using an LLM. We find that training on code-switched data aligns the representations of parallel text, particularly across different scripts, and that this alignment persists through training on monolingual documents. Under a learning curriculum that progresses from word-level code-switching, to sentence-level code-switching, to monolingual documents, models trained on code-switched data outperform baselines trained without it on the BabyLM evaluation suite. Our work characterizes code-switching curriculum learning as an effective data augmentation method for multilingual pretraining. We release our code, data, and models at https://github.com/drooryck/multilingual-macaroni.

Voice agents are increasingly deployed in workflows where failed interactions can affect transactions, access, and other consequential outcomes, creating a need for reproducible and interpretable evaluation. We introduce Inquesto Score (IS), a protocol for measuring voice-agent reliability as the percentage of calls in a fixed, versioned evaluation population that achieve the caller's goal without a functional failure or worse. Rather than combining heterogeneous metrics, IS defines explicit failure events and severity levels and evaluates the deployed voice pipeline. Timing failures, including talk-over and delayed responses, are measured directly from audio, while semantic and state-dependent failures are evaluated using scenario predicates, tool traces, and a pinned open-model judge. Diagnostic views of behavior, acoustic robustness, identity handling, and speaker groups accompany the score without being combined into it. Inquesto Score v0.1 evaluates 30 scenarios, three acoustic conditions, four speaker groups, and 306 calls per agent across 13 configurations of a reference voice-agent system. Our evaluation shows that reliable measurement requires evidence beyond transcripts, explicit treatment of deployment conditions, and validation of the evaluators used to determine outcomes. We release the protocol, reference implementation, and evaluation records.

Language models often produce homogeneous responses to open-ended tasks; such homogeneity can spawn groupthink-the convergence of ideas toward a singular and potentially suboptimal decision. We formulate persona diversification as a set-level conditioning problem and study two orthogonal design choices: selecting versus generating personas, and space-filling versus frontier-seeking diversity. We instantiate this design space with four methods spanning coverage and dispersion subset selections, uniform-coverage sampling, and evolutionary persona generation. Evaluations on the Alternative Uses Task (AUT), Infinity-Chat, and Divergent Association Task (DAT) show the benefits of the proposed methods across tasks and creativity objectives. On AUT, evolutionary persona generation increases response diversity by 78.8%, originality by 26.1%, flexibility by 49.5%, and holistic creativity by 13.9% over task-only prompting, while maintaining 98.5% validity; on Infinity-Chat, it nearly doubles persona-induced response separation relative to random personas. Moreover, evolutionary personas compose with creativity-optimized prompting, further increasing its response diversity by 18.6% and creativity by 6.3%. These results establish persona-set geometry as a task-agnostic mechanism for eliciting divergent LLM outputs, and support persona diversification as a reusable complement to prompt optimization.

Audio language models state numbers for acoustic quantities, and neither human opinion nor a judge model says whether such a number is true of the signal. AcoustiClaim extracts each numeric claim from free text, scores it against the instrument that defines the quantity, and classes each quantity by where its reference can be read. Four open-weight systems and one closed model, asked for ten quantities five ways on two corpora, fill 207 cells. Of these, 49 emit fewer than five distinct values, and eight of the 158 cells that can be ranked exceed a rank correlation of 0.3, the bar we set, three with an interval clear of it, five of them one closed model reading pitch. Error sits at or above a constant-predictor floor in every ranked cell but three. The reference decoder we train declines the five voice quantities in prose on 95% of mixtures, with nothing withheld, and states them on the clean twins, reproducing its targets' rule from audio alone. With a calibrated threshold, withholding lowers error on all ten quantities on the mixtures in the mean and on eight at every split, against at most 0.6% from a random selector. A linear baseline orders errors at least as well as ours. F0 s.d. and shimmer stay above the constant floor.

Target-speaker automatic speech recognition (TS-ASR) must identify and transcribe a desired speaker under varying overlap and noise conditions. These changes alter the acoustic evidence for the target speaker in the speech mixture, motivating inference-time calibration of speaker conditioning. We introduce asymmetric classifier-free guidance (CFG) for TS-ASR using Whisper: the speaker-conditioned branch predicts the target transcript, while the speaker-unconditioned branch predicts serialized multi-speaker transcripts. CFG adjusts the contribution of speaker conditioning during decoding through a single guidance scale. We select a global guidance scale on target-domain development data and train a lightweight encoder-based predictor to adjust it for each utterance, keeping the recognition model fixed. Under domain shifts, our full system achieves relative word error rate (WER) reductions of up to 21.8% over the condition-only baseline, and 5.6% over standard conditional decoding of the same CFG-trained model. Oracle analysis shows that substantially larger WER reductions are possible through utterance-level scale selection and identifies how beneficial adjustments vary with domain shifts.

Reference-based LLM-as-a-judge evaluation assumes the reference answer is the target. In deployed agentic systems that operate over dynamic entities (support cases, assets, accounts), the closest available reference typically applies the correct procedure to a different entity, so a literal judge penalizes different identifiers, dates, and statuses as errors or hallucinations. We name this failure mode reference-instance divergence (RID). We propose CARGO, a framework that (i) treats retrieved references as procedural exemplars and grounds factual judgments in the live instance's observed context, (ii) assigns each claim a three-way status (supported, contradicted, unverifiable) and penalizes only contradictions, and (iii) gates evaluation by retrieval confidence, casting production evaluation as selective prediction. We introduce CARGO-Bench, a perturbation-based diagnostic suite with ground truth by construction that separates leniency from discrimination. On CARGO-Bench (246 items, two judge models, 7,872 judgments), the standard reference-based judge penalizes 100% of correct entity-transplanted answers and is uninformative (discrimination index DI ~ 0); supplying the live facts without reframing changes nothing. CARGO eliminates these false penalties (0/50) while retaining near-complete contradiction recall (50/50 and 49/50), raising DI to 0.58 [0.48, 0.68]; a rubric-swap control attributes most of the effect to context-grounded dimension definitions. CARGO also exposes a limitation of its own design: the leniency that protects entity values suppresses detection of procedural corruptions (20% recall). A post-hoc fix does not close the gap, and an LLM-as-annotator study with written guidelines and adjudication shows the same blind spot. We release a preregistered protocol for extending the evaluation to expert agreement, risk-coverage, and cost on production traffic.

Retrieval-based factuality evaluation, where LLM-generated claims are verified against evidence from authoritative medical corpora, has become the dominant paradigm for scalable hallucination detection in high-stakes clinical settings. Despite the urgency of reliable and transparent medical fact verification, most systems measure performance with aggregate metrics like F1, which obscure where and why failures occur. Existing RAG diagnostics require gold answers or annotated gold evidence, neither of which exists in this regime. We introduce two comprehensive taxonomies, grounded in a case study on the open-ended MedExpert dataset and 3 closed-ended datasets, decomposing failures into retrieval-stage errors along five quality dimensions, and verifier-reasoning errors into six consecutive steps. We adapt an automatic pattern induction pipeline using LLM-as-Judge to label evidence quality and classify verifier reasoning errors at scale, and then stress-test our findings across 4 retrieval methods and 6 frontier verifier models. Our analysis reveals that scaling model size, adding reasoning effort, expanding to authoritative web sources, and applying medical fine-tuning do not resolve these failure modes, demonstrating that they represent fundamental limitations of the retrieve-then-verify paradigm in open-ended medical settings rather than artifacts of outdated systems. We release our code and data at https://anonymous.4open.science/r/Medical_RAG_eval-4AB5 for the full reproducibility of our results.

Mixture-of-experts (MoE) models activate only a few experts per token but store the entire expert pool. Pruning this pool requires identifying experts whose removal preserves model behavior. Routing frequency and output magnitude do not fully describe deletion damage, which also depends on how the surviving and replacement experts compensate for the removed output. We introduce RAZOR, a training-free pruning method based on consensus residuals, the deviations of expert outputs from their original weighted mixture. At a fixed layer input, these residuals give the exact output change for a single deletion under survivor renormalization and router refill. RAZOR aggregates this damage by conditional root mean square and selects experts under a layerwise budget using forward computation alone, without gradients, subset search, or recovery training. Against frequency, activation-norm, and REAP baselines on GLM-4.7-Flash and Qwen3.6-35B-A3B at 25% and 50% expert removal, it attains the highest macro average over nine reasoning-intensive tasks in all four model-budget settings, gaining 2.12-5.59 points over REAP and lowering reverse KL in all four. On DeepSeek-V4-Flash-0731 and Hy3, it also achieves the highest macro average among the three residual criteria. Local exactness does not guarantee better joint pruning. Generation analyses show changes in diversity, formatting, and termination despite higher task scores.

We introduce target-speaker unlearning ASR (TSU-ASR) task in a fully end-to-end framework for multi-speaker ASR and diarization. Given a multi-speaker utterance and a set of opt-out speakers who do not wish to have their speech transcribed, the task requires an ASR system to transcribe all speakers except the opt-out ones, while still indicating when those speakers are active. As a first step towards tackling this task, we introduce a novel, light-weight Enrollment-Conditioned Gating (ECG) module attachable to a frozen dual-stream speech LLM that enables ASR for new opt-out speakers dynamically during inference, even those who were not seen during initial ECG training phase. Our experiments on both AMI (English) and AliMeeting (Mandarin) datasets show that speech transcription accuracy for corresponding opt-out words or characters falls from 72.3% to 48.2% and from 73.6% to 27.3%, respectively, while retained speakers' transcription error rates maintain more or less the same. Our approach provides a practical solution for modern video conferencing platforms, allowing speakers to dynamically opt-out from automated AI transcriptions without forcefully leaving the meeting sessions, enabling a privacy-preserving interface for potentially millions of online meetings daily.

Large Language Models (LLMs) have shown strong performance in low-resource offline translation; however, extending them to simultaneous speech-to-speech translation (Simul-S2ST) remains challenging due to the scarcity of causally aligned training data with high cross-lingual speaker fidelity. In addition, existing approaches rely on fixed translation policy or confidence heuristics, leading to suboptimal quality and higher latency. We propose a causality-aware Simul-S2ST framework with a novel data pipeline that generates high-fidelity, causally aligned segments with improved voice transfer. The framework introduces (i) a factorized S2ST architecture (FAST), (ii) a causality-aware adaptive policy (CAP), and (iii) causality-aware latency metric. Experiments on CVSS Spanish, German, and French show that FAST-CAP consistently improves the quality-latency trade-off, achieving up to +1.2 BLEU and a 26% relative latency reduction over a fixed policy. Despite using substantially less training data than existing systems, FAST-CAP achieves state-of-the-art results in speech translation quality and speaker fidelity while yielding up to a 38.8% relative reduction in latency.

Cognitive psychology has studied how people encode, use, and learn concepts that describe categories, and how they represent, recognize, and acquire chunks for familiar patterns of elements. The literatures on these two topics are nearly disjoint, which poses a challenge for unified theories of cognition. In this paper, we review Cobweb, a computational account of categorization and concept formation, then propose an extended theory that incorporates chunks and their acquisition. The theory makes no commitments about modality, applying to any experience that decomposes into elements and relations among them. We also present \trellis/, an implementation of this theory, and illustrate its application to learning context-free grammars, which we adopt as a testbed because they involve both concept-like and chunk-like elements. In addition, we report experimental results on learning for three synthetic grammars that demonstrate the system's ability to represent syntactic knowledge, use it to parse and generate sentences, and learn compositional structures from sample parses. We conclude by discussing related work on concepts and chunks, along with directions for future research on the problem.