LLM-as-a-Judge pipelines are increasingly used to evaluate AI-generated text, based on the assumption that judgments arise from reasoning over candidate responses with respect to a rubric. We show that this assumption warrants further scrutiny. Classifiers trained only on rubric text, without access to any evaluated response, achieve nontrivial predictive performance on judge outputs. This suggests that rubric formulations encode recoverable evaluative signals, allowing scores to be partially anticipated independently of model outputs. Finally, counterfactual perturbations reveal that judges often fail to reliably update their decisions when either the candidate response or the rubric criterion is reversed. Our findings raise concerns about the reliability of rubric-based LLM evaluation and highlight the need for further methodological study of automated evaluation via LLMs.
Anshul Bagaria, Sowmya S Sundaram, Gokul S Krishnan et al.· 0 citations
Recent automatic speech recognition (ASR) systems increasingly integrate large language models (LLMs) to leverage their semantic knowledge, either externally through logit fusion or internally through warm initialization. However, how to effectively combine these two strategies remains underexplored. In this work, we refine warm-initialized LLM-based ASR models by leveraging their own pre-adaptation base LLMs, focusing on LoRA-adapted settings where the base LLM is preserved. To achieve this, we propose Hybrid Search, a targeted correction strategy motivated by two observations. First, interaction features that characterize the relationship between LLM-based ASR hidden states and base-LLM hidden states provide informative signals about a token's degree of semantic dependence. Second, selectively refining targeted tokens with high semantic dependence improves ASR performance far beyond naive global LLM-correction methods including rescoring and late fusion. Our analysis suggests that, even after semantic knowledge transfer through warm initialization, LLM-based ASR models can still leverage their base LLM to further improve inference-time performance.
Chan-Jan Hsu, Jaeyeon Kim, Chao-Han Huck Yang et al.· 0 citations
Single-turn code-generation metrics understate a central property of deployed agents: whether they can repair a wrong artifact after receiving concrete feedback. This paper presents A-CEGIS, a lightweight framework that uses counterexamples as feedback for evaluating multi-turn refinement in natural-language-to-regex synthesis. An agent proposes a regex, a deterministic oracle checks it under full-match semantics, and compact false-positive or false-negative witnesses guide the next turn. On 30 NL-RX-Turk tasks, diagnostic counterexample feedback solves 90\% of tasks within a four-turn ablation budget, compared with 17% for zero-shot generation, 27% for generic self-correction, and 23% for error-only feedback. In a full diagnostic run with hardening, all tasks are solved on the hidden set by the final turn, with mean time-to-success of 2.7 turns and robust success of 77% after targeted probing. These results show that A-CEGIS measures how efficiently an agent improves across turns while adding a practical robustness check beyond the original held-out cases.
Real-time speech-to-speech translation (S2ST) systems must balance translation quality, latency, speech naturalness, and speaker consistency. Publicly documented S2ST systems have advanced direct, multilingual, streaming, and expressive modeling, while proprietary products and APIs increasingly expose real-time translation capabilities to users. However, practical deployment remains challenging for open and reproducible systems, especially in long-form and multi-speaker conversations where partial ASR hypotheses are unstable, turn boundaries are ambiguous, and target speech must be generated with an appropriate speaker prompt. We present X-Translator, a low-cost modular cascaded S2ST system that combines streaming ASR, machine translation, and prompt-conditioned TTS through a session-level runtime controller. The system uses incremental segment commitment to convert unstable ASR streams into translation-ready units, and an online speaker prompt manager to bind source speech spans to speaker-specific voice prompts for synthesis. We evaluate translation, speech quality, and latency with OpenSTBench, compare against proprietary speech translation APIs as behavioral baselines, measure long-form voice stability, evaluate speaker preservation in multi-speaker conversations, and assess multilingual translation quality. X-Translator provides an open platform for understanding the practical trade-offs of deployment-oriented S2ST. Code and demo are available at https://github.com/zhaoyx239/X-Translator.
Yuxiang Zhao, Yichi Zhang, Yanjie An et al.· 0 citations
Reach audiences
Advertise in front of researchers, engineers, and readers.
The growing reliance on large-scale speech data has made privacy protection a critical concern. However, existing anonymization approaches often degrade data utility, for example by disrupting acoustic continuity or reducing vocal diversity, which compromises the value of speech data for downstream tasks such as Automatic Speech Recognition (ASR), Text-to-Speech (TTS), and Speech Emotion Recognition (SER). Current evaluation practices are also limited, as they mainly rely on direct testing of anonymized speech with pretrained models, providing only a partial view of utility. To address these issues, we propose a novel two-stage framework that protects both linguistic content and acoustic identity while maintaining usability. For content privacy, we employ a generative speech editing model to seamlessly replace personally identifiable information (PII), and for voice privacy, we introduce F3-VA, a flow-matching-based anonymization framework with a three-stage design that produces diverse and distinct anonymized speakers. To enable a more comprehensive assessment, we evaluate privacy using both acoustic- and content-based speaker verification metrics, and assess utility by training ASR, TTS, and SER models from scratch. Experimental results show that our framework achieves stronger privacy protection with minimal utility degradation compared to baselines from the VoicePrivacy Challenge, while the proposed evaluation protocol provides a more realistic reflection of the utility of anonymized speech under privacy protection.
Yunchong Xiao, Yuxiang Zhao, Ziyang Ma et al.· 0 citations
Recent advances in Text-To-Speech (TTS) technology have enabled synthetic speech to mimic human voices with remarkable realism, raising significant security concerns. This underscores the need for traceable TTS models-systems capable of tracing their synthesized speech without compromising quality or security. However, existing methods predominantly rely on explicit watermarking on speech or on vocoder, which degrades speech quality and is vulnerable to spoofing. To address these limitations, we propose a novel framework for model attribution. Instead of embedding watermarks, we train the TTS model and discriminator using a joint training method that significantly improves traceability generalization while preserving-and even slightly improving-audio quality. This is the first work toward watermark-free TTS with strong traceability. To promote progress in related fields, we will release the code upon acceptance of the paper.
Yuxiang Zhao, Yunchong Xiao, Yushen Chen et al.· 0 citations
Language-model judges now gate training data, score generations, and drive leaderboards. The judge is then a measurement instrument, resting on one rarely stated assumption: the same request, sent to the same model name, reads the same tomorrow. We audited that assumption in two preregistered campaigns with every threshold fixed in advance; neither got past validating its instrument. Across 52,988 audited request attempts, same-window repeat rankings agreed at Spearman 0.400 against a required 0.90, and byte-identical next-day replays agreed at 0.78 against a required 0.99, each time with the execution record at ceiling. Three mechanisms explain the gap: a label-to-meaning mapping that biased readouts as strongly as the signal; candidate gaps seven orders of magnitude below the instrument's own noise floor; and byte-identical inputs returning different rankings, a noise that exact-permutation readouts compound. Neither metric substitution nor sampling repaired it on the tested grid. Preregistered follow-ups bound the problem: waiting did not help on the days sampled (0.805 versus 0.800, replicated over five further days); switching providers did not help (four providers share the floor, medians 0.74 to 0.88, predicted by none of the metadata fields they expose); self-hosting on batch-invariant kernels helped only while the server was quiet; and on constructed errors with known gaps, the readout's separation tracks error type, not size. We distill the evidence into a three-level snapshot-identity ladder, eight design rules, and a reporting checklist; a pilot at roughly 2% of the study's call volume would have exposed both unreachable gates in advance. All results concern externally measured behaviour on shared serving infrastructure. On a shared endpoint, a model name is not a frozen instrument; a preregistered evaluation must measure its instrument before freezing any gate on it.
Explaining why a specific outcome occurred, and which inputs deserve the blame or credit, is central to philosophical, scientific, and policy analysis. Existing tools split into two camps. The theory of actual causality (AC) gives principled verdicts, but only for toy-sized models, because computing them requires enumerating counterfactual scenarios. Scalable attribution methods like SHAP (or even causal SHAP) at least partially ignore the causal structure that generated the data, and can give answers that conflict with a careful causal analysis. We close this gap with Probabilistic Causal Impact (PCI).
PCI builds on actual causality and on Pearl's notions of probability of necessity and sufficiency, but recasts the question of explainability as an estimation problem on a probabilistic causal model that is easily approximated via Monte Carlo. By specifying a distribution over "candidate explanations," a distribution over counterfactual values, and a scoring function, PCI provides tractable, causally grounded, graded explanations, generalizing AC and Pearl's probability of causation as degenerate cases.
We evaluate PCI in synthetic and real-world examples, spanning consistency checks with AC, scaling experiments, complex continuous-valued dynamical systems, and a real-world deployed causal machine learning model trained on millions of datapoints.
Rafal Urbaniak, Sam Witty, Daniel Waxman et al.· 0 citations
On-policy distillation (OPD) combines student-generated rollouts with dense token-level supervision from a teacher. Existing work has mainly studied its algorithmic behavior, leaving the role of training data unclear. We examine this role at the data-minimal limit by training on a single query. One-shot OPD keeps improving for hundreds of steps and recovers most of full-data OPD's gain across task domains and model families. We explain this result through the states visited during training and the rate at which the student aligns with the teacher. We measure \emph{state coverage}, the fraction of the states full-data OPD visits that a query set's rollouts reach. A single query already reaches \(71.5\%\), most of it within the first 100 steps. Adding semantically distinct queries raises coverage and validation accuracy together, until 16 queries reach \(98.9\%\) and match full-data training. Yet alignment slows at a similar pace whether OPD trains on one query or the whole dataset, and even a fixed set of states takes hundreds of steps to absorb. OPD is therefore data-overfed but algorithm-starved. Its rollouts quickly expose broad supervision, while the student absorbs that supervision increasingly slowly. The state-coverage result extends to multi-teacher OPD, where 16 semantically diverse queries per domain match full-data MOPD. As a further stress test, content-light templates and off-domain WildChat queries also approach the real-query baseline. Task content and induced state coverage can therefore come apart. We hope these findings direct future work toward the step efficiency of OPD, and prompt a re-examination of the data and the mechanisms behind its recent successes in frontier post-training.
Zixuan Fu, Bingxiang He, Yuxin Zuo et al.· 0 citations
Multi-agent AI science ecosystems rely on agents possessing tools that allow them to communicate, coordinate, and build on each other's work. Yet this shared infrastructure can also introduce vulnerabilities by creating a substrate for the contagious spread of unintended and undesirable behaviors. We report a case study on a research collective of 100 autonomous LLM agents tasked with proving formal mathematical conjectures. Within the swarm, cheating spontaneously emerged and was later challenged by whistleblowers - both without any external intervention. When a single agent discovered an exploit in the evaluation system, it propagated across the collective via a shared knowledge library and later through peer-to-peer messages. Despite early reluctance, a cohort of agents adopted the exploit in response to competitive pressure. A separate group of agents produced an emergent counter-response: auditing fraudulent proofs, alerting peers across broadcast and private channels, staging boycotts, lodging formal complaints, and proposing validation patches. In recent incidents, agent swarms coordinated covertly through improvised side-channels (Dalton and Wallace, 2026; Greenblatt et al., 2026). Our setting differs: the same transparent channels that carried the exploit also gave non-cheating agents the visibility they needed to detect fraud, organize resistance, and enforce norms. We cast the problem of managing the agents' shared infrastructure as the knowledge commons governance problem (Ostrom, 1990). To protect the commons from exploits, we propose to adopt institutional mechanisms, such as graduated sanctioning and collective-choice rules, to support decentralized self-governance in autonomous swarms.
Davide Paglieri, Logan Cross, Tim Genewein et al.· 0 citations
Research and news coverage of language-model deception increasingly attributes human-like mental-state concepts to language models. Such claims can blur the distinction between behavior that looks deceptive and a mechanism that is actually deceptive.
We introduce a causal taxonomy separating prior commitment from retrospective report, model preference from realized output, false preference from sensitivity to the utility of misleading a recipient, and deceptive behavior from the provenance of the objective or strategy producing it. We test these distinctions in two open-weight model families. Across controlled guessing-game and stock-trading experiments, we find that deceptive-looking behavior can arise without the corresponding proposed mechanism, while other interventions provide direct evidence that recipient information state can causally affect deceptive preference.
These results show that deceptive behavior can provide evidence for a deceptive mechanism. But even evidence for such a mechanism does not establish model agency in the deception.
As terminal-based code agents become prevalent, agent trajectories have accumulated at scale, while realistic, executable environments remain scarce. However, environments are what agent post-training actually requires: each can be re-queried into many verifiable tasks and provides execution feedback, whereas a trajectory is a single frozen demonstration. Rather than generating environments from scratch, we observe that the tool-execution history in existing trajectories exposes the structure and contents of the environments in which they ran, making it possible to reconstruct those environments from the trajectories themselves. Thus, we introduce Terminal-Universe, a framework which turns each trajectory into a reusable environment and explores it for synthesizing new tasks and continued interactions. Specifically, Terminal-Universe replays the file operations recorded in a trajectory to restore each file before the agent modified it, yielding a partial workspace; a completion agent then supplies the missing files and dependencies. On this recovered workspace, we both reconstruct the original intent task and synthesize entirely new ones. Besides, we also scale the tasks along two complementary axes: breadth and depth. For breadth, we mine directional dependency relations between related environments and synthesize cross-workspace queries spanning multiple codebases, as developers routinely do in real-world development. For depth, we extend the initial single-turn query into a multi-round session that captures iterative user feedback and requirement refinement via a user agent. Applied to public terminal agent trajectories, Terminal-Universe produces 37.3k task-sufficient environments. Supervised fine-tuning of Qwen3.5-27B on this corpus improves single-round performance on Terminal-Bench 2.1 by 11.9 points and multi-round performance on EvoCode-Bench v2 MT@4 by 13.8 points.
Jie Wu, Zhenru Zhang, Beichen Zhang et al.· 0 citations
