PodEval: A Multimodal Evaluation Framework for Podcast Audio Generation

Recently, an increasing number of multimodal (text and audio) benchmarks have emerged, primarily focusing on evaluating models' understanding capability. However, exploration into assessing generative capabilities remains limited, especially for open-ended long-form content generation. Significant challenges lie in no reference standard answer, no unified evaluation metrics and uncontrollable human judgments. In this work, we take podcast-like audio generation as a starting point and propose PodEval, a comprehensive and well-designed open-source evaluation framework. In this framework: 1) We construct a real-world podcast dataset spanning diverse topics, serving as a reference for human-level creative quality. 2) We introduce a multimodal evaluation strategy and decompose the complex task into three dimensions: text, speech and audio, with different evaluation emphasis on "Content" and "Format". 3) For each modality, we design corresponding evaluation methods, involving both objective metrics and subjective listening test. We leverage representative podcast generation systems (including open-source, close-source, and human-made) in our experiments. The results offer in-depth analysis and insights into podcast generation, demonstrating the effectiveness of PodEval in evaluating open-ended long-form audio. This project is open-source to facilitate public use: https://github.com/yujxx/PodEval.

BatonVoice: An Operationalist Framework for Enhancing Controllable Speech Synthesis with Linguistic Intelligence from LLMs

The rise of Large Language Models (LLMs) is reshaping multimodel models, with speech synthesis being a prominent application. However, existing approaches often underutilize the linguistic intelligence of these models, typically failing to leverage their powerful instruction-following capabilities. This limitation hinders the model's ability to follow text instructions for controllable Text-to-Speech~(TTS). To address this, we propose a new paradigm inspired by ``operationalism'' that decouples instruction understanding from speech generation. We introduce BatonVoice, a framework where an LLM acts as a ``conductor'', understanding user instructions and generating a textual ``plan'' -- explicit vocal features (e.g., pitch, energy). A separate TTS model, the ``orchestra'', then generates the speech from these features. To realize this component, we develop BatonTTS, a TTS model trained specifically for this task. Our experiments demonstrate that BatonVoice achieves strong performance in controllable and emotional speech synthesis, outperforming strong open- and closed-source baselines. Notably, our approach enables remarkable zero-shot cross-lingual generalization, accurately applying feature control abilities to languages unseen during post-training. This demonstrates that objectifying speech into textual vocal features can more effectively unlock the linguistic intelligence of LLMs.

The Dragon Hatchling: The Missing Link between the Transformer and Models of the Brain

The relationship between computing systems and the brain has served as motivation for pioneering theoreticians since John von Neumann and Alan Turing. Uniform, scale-free biological networks, such as the brain, have powerful properties, including generalizing over time, which is the main barrier for Machine Learning on the path to Universal Reasoning Models. We introduce `Dragon Hatchling' (BDH), a new Large Language Model architecture based on a scale-free biologically inspired network of \$n\$ locally-interacting neuron particles. BDH couples strong theoretical foundations and inherent interpretability without sacrificing Transformer-like performance. BDH is a practical, performant state-of-the-art attention-based state space sequence learning architecture. In addition to being a graph model, BDH admits a GPU-friendly formulation. It exhibits Transformer-like scaling laws: empirically BDH rivals GPT2 performance on language and translation tasks, at the same number of parameters (10M to 1B), for the same training data. BDH can be represented as a brain model. The working memory of BDH during inference entirely relies on synaptic plasticity with Hebbian learning using spiking neurons. We confirm empirically that specific, individual synapses strengthen connection whenever BDH hears or reasons about a specific concept while processing language inputs. The neuron interaction network of BDH is a graph of high modularity with heavy-tailed degree distribution. The BDH model is biologically plausible, explaining one possible mechanism which human neurons could use to achieve speech. BDH is designed for interpretability. Activation vectors of BDH are sparse and positive. We demonstrate monosemanticity in BDH on language tasks. Interpretability of state, which goes beyond interpretability of neurons and model parameters, is an inherent feature of the BDH architecture.

Voice Evaluation of Reasoning Ability: Diagnosing the Modality-Induced Performance Gap

We present Voice Evaluation of Reasoning Ability (VERA), a benchmark for evaluating reasoning ability in voice-interactive systems under real-time conversational constraints. VERA comprises 2,931 voice-native episodes derived from established text benchmarks and organized into five tracks (Math, Web, Science, Long-Context, Factual). Each item is adapted for speech interaction while preserving reasoning difficulty. VERA enables direct text-voice comparison within model families and supports analysis of how architectural choices affect reliability. We assess 12 contemporary voice systems alongside strong text baselines and observe large, consistent modality gaps: on competition mathematics a leading text model attains 74.8% accuracy while its voice counterpart reaches 6.1%; macro-averaged across tracks the best text models achieve 54.0% versus 11.3% for voice. Latency-accuracy analyses reveal a low-latency plateau, where fast voice systems cluster around ~10% accuracy, while approaching text performance requires sacrificing real-time interaction. Diagnostic experiments indicate that common mitigations are insufficient. Increasing "thinking time" yields negligible gains; a decoupled cascade that separates reasoning from narration improves accuracy but still falls well short of text and introduces characteristic grounding/consistency errors. Failure analyses further show distinct error signatures across native streaming, end-to-end, and cascade designs. VERA provides a reproducible testbed and targeted diagnostics for architectures that decouple thinking from speaking, offering a principled way to measure progress toward real-time voice assistants that are both fluent and reliably reasoned.

Convergence and Divergence of Language Models under Different Random Seeds

In this paper, we investigate the convergence of language models (LMs) trained under different random seeds, measuring convergence as the expected per-token Kullback--Leibler (KL) divergence across seeds. By comparing LM convergence as a function of model size and training checkpoint, we identify a four-phase convergence pattern: (i) an initial uniform phase, (ii) a sharp-convergence phase, (iii) a sharp-divergence phase, and (iv) a slow-reconvergence phase. Further, we observe that larger models reconverge faster in later training stages, while smaller models never actually reconverge; these results suggest that a certain model size may be necessary to learn stable distributions. Restricting our analysis to specific token frequencies or part-of-speech (PoS) tags further reveals that convergence is uneven across linguistic categories: frequent tokens and function words converge faster and more reliably than their counterparts (infrequent tokens and content words). Overall, our findings highlight factors that influence the stability of the learned distributions in model training.

Scaling Spoken Language Models with Syllabic Speech Tokenization

Spoken language models (SLMs) typically discretize speech into high-frame-rate tokens extracted from SSL speech models. As the most successful LMs are based on the Transformer architecture, processing these long token streams with self-attention is expensive, as attention scales quadratically with sequence length. A recent SSL work introduces acoustic tokenization of speech at the syllable level, which is more interpretable and potentially more scalable with significant compression in token lengths (4-5 Hz). Yet, their value for spoken language modeling is not yet fully explored. We present the first systematic study of syllabic tokenization for spoken language modeling, evaluating models on a suite of SLU benchmarks while varying training data scale. Syllabic tokens can match or surpass the previous high-frame rate tokens while significantly cutting training and inference costs, achieving more than a 2x reduction in training time and a 5x reduction in FLOPs. Our findings highlight syllable-level language modeling as a promising path to efficient long-context spoken language models.

The Unheard Alternative: Contrastive Explanations for Speech-to-Text Models

Contrastive explanations, which indicate why an AI system produced one output (the target) instead of another (the foil), are widely regarded in explainable AI as more informative and interpretable than standard explanations. However, obtaining such explanations for speech-to-text (S2T) generative models remains an open challenge. Drawing from feature attribution techniques, we propose the first method to obtain contrastive explanations in S2T by analyzing how parts of the input spectrogram influence the choice between alternative outputs. Through a case study on gender assignment in speech translation, we show that our method accurately identifies the audio features that drive the selection of one gender over another. By extending the scope of contrastive explanations to S2T, our work provides a foundation for better understanding S2T models.

Conversational Implicatures: Modelling Relevance Theory Probabilistically

Recent advances in Bayesian probability theory and its application to cognitive science in combination with the development of a new generation of computational tools and methods for probabilistic computation have led to a 'probabilistic turn' in pragmatics and semantics. In particular, the framework of Rational Speech Act theory has been developed to model broadly Gricean accounts of pragmatic phenomena in Bayesian terms, starting with fairly simple reference games and covering ever more complex communicative exchanges such as verbal syllogistic reasoning. This paper explores in which way a similar Bayesian approach might be applied to relevance-theoretic pragmatics (Sperber & Wilson, 1995) by study a paradigmatic pragmatic phenomenon: the communication of implicit meaning by ways of (conversational) implicatures.

Bridging Fairness and Explainability: Can Input-Based Explanations Promote Fairness in Hate Speech Detection?

Natural language processing (NLP) models often replicate or amplify social bias from training data, raising concerns about fairness. At the same time, their black-box nature makes it difficult for users to recognize biased predictions and for developers to effectively mitigate them. While some studies suggest that input-based explanations can help detect and mitigate bias, others question their reliability in ensuring fairness. Existing research on explainability in fair NLP has been predominantly qualitative, with limited large-scale quantitative analysis. In this work, we conduct the first systematic study of the relationship between explainability and fairness in hate speech detection, focusing on both encoder- and decoder-only models. We examine three key dimensions: (1) identifying biased predictions, (2) selecting fair models, and (3) mitigating bias during model training. Our findings show that input-based explanations can effectively detect biased predictions and serve as useful supervision for reducing bias during training, but they are unreliable for selecting fair models among candidates.

FLEXI: Benchmarking Full-duplex Human-LLM Speech Interaction

Full-Duplex Speech-to-Speech Large Language Models (LLMs) are foundational to natural human-computer interaction, enabling real-time spoken dialogue systems. However, benchmarking and modeling these models remains a fundamental challenge. We introduce FLEXI, the first benchmark for full-duplex LLM-human spoken interaction that explicitly incorporates model interruption in emergency scenarios. FLEXI systematically evaluates the latency, quality, and conversational effectiveness of real-time dialogue through six diverse human-LLM interaction scenarios, revealing significant gaps between open source and commercial models in emergency awareness, turn terminating, and interaction latency. Finally, we suggest that next token-pair prediction offers a promising path toward achieving truly seamless and human-like full-duplex interaction.