Paper deep dive
Listen, Reason, and Segment: Aligning LALMs with Editorial Judgment for Media Chapterization
Tony Alex, Wish Suharitdamrong, Sara Atito, Armin Mustafa, Muhammad Awais, Philip J. B. Jackson, Jiankang Deng, Ismail Elezi
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 95%
Last extracted: 8/23/2026, 2:20:12 AM
Summary
This paper introduces AudioChaps, a post-training framework for aligning Large Audio Language Models (LALMs) for automated audio chapterization. The authors address the gap in practical media workflows by treating chapterization as a subjective editorial judgment task rather than just acoustic event detection. They propose a two-stage alignment process using Group Relative Policy Optimization (GRPO) guided by Chain-of-Thought (CoT) reasoning. The framework includes three curated datasets: AudioChaps-Alignment, AudioChaps-CoT, and AudioChaps-Eval. The resulting model, AudioChaps-R1, significantly outperforms state-of-the-art baselines like Audio-Flamingo-3-Think, demonstrating that GRPO-trained LALMs can reliably segment audio into thematically coherent chapters.
Entities (14)
Relation Signals (9)
AudioChaps-Alignment → derivedfrom → YouTube
confidence 97% · AudioChaps-Alignment, derived from creator-annotated chapter boundaries on YouTube
AudioChaps-R1 → outperforms → Audio-Flamingo-3-Think
confidence 96% · The AudioChaps framework produces our final aligned LALM, AudioChaps-R1, which improves average F1 by 49 points [over Audio-Flamingo-3-Think].
AudioChaps-Eval → usedfor → Audio Chapterization
confidence 95% · AudioChaps-Eval, a held-out benchmark for audio chapterization.
AudioChaps → uses → SFT
confidence 95% · AudioChaps-R1 is trained in two stages: SFT on the AudioChaps-CoT corpus establishes a structured, evidence-grounded reasoning prior, after which GRPO calibrates boundary decisions toward creator-authored annotations.
AudioChaps → uses → GRPO
confidence 95% · We present AudioChaps, a post-training framework for aligning end-to-end LALMs for this task via Group Relative Policy Optimization (GRPO) guided by Chain-of-Thought (CoT) reasoning.
AudioChaps-R1-Zero → outperforms → Audio-Flamingo-3-Think
confidence 94% · Applying GRPO directly without a Supervised Fine-Tuning (SFT) cold start, AudioChaps-R1-Zero already improves average F1 by 33 points over the state-of-the-art LALM Audio-Flamingo-3-Think.
AudioChaps-CoT → usedfor → SFT
confidence 93% · SFT on the AudioChaps-CoT corpus establishes a structured, evidence-grounded reasoning prior
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Large Audio Language Models (LALMs) have made rapid progress on standardized benchmarks, yet their deployment in practical media workflows, curation, archival indexing, and content distribution remains largely unrealized. We identify automated audio chapterization, the task of segmenting continuous audio streams into thematically coherent chapters, as a demanding and commercially consequential setting that exposes this gap. Chapterization is challenging because boundaries are defined less by objective acoustic events than by subjective editorial judgment, requiring models to reason sequentially over long acoustic contexts and approximate creator-authored boundary decisions. We present AudioChaps, a post-training framework for aligning end-to-end LALMs for this task via Group Relative Policy Optimization (GRPO) guided by Chain-of-Thought (CoT) reasoning. To support training and evaluation, we curate three datasets: AudioChaps-Alignment, derived from creator-annotated chapter boundaries on YouTube; AudioChaps-CoT, which provides structured supervision for well-formatted, high-quality, and evidence-grounded boundary reasoning; and AudioChaps-Eval, a held-out benchmark for audio chapterization. Applying GRPO directly without a Supervised Fine-Tuning (SFT) cold start, AudioChaps-R1-Zero already improves average F1 by 33 points over the state-of-the-art LALM Audio-Flamingo-3-Think. The AudioChaps framework produces our final aligned LALM, AudioChaps-R1, which improves average F1 by 49 points. These results demonstrate that GRPO-trained LALMs can reliably transform unstructured auditory streams into navigable, structured media. Our code, models, and dataset resources will be released upon acceptance at this https URL.
Tags
Links
- Source: https://arxiv.org/abs/2608.16539v1
- Canonical: https://arxiv.org/abs/2608.16539v1
Trouble viewing inline? Open PDF directly →
Full Text
73,417 characters extracted from source content.
Expand or collapse full text
Listen, Reason, and Segment: Aligning LALMs with Editorial Judgment for Media Chapterization Tony Alex * † University of Surrey Wish Suharitdamrong University of Surrey Sara Atito University of Surrey Armin Mustafa University of Surrey Muhammad Awais University of Surrey Philip J. B. Jackson University of Surrey Jiankang Deng * Huawei Noah’s Ark Lab Ismail Elezi Huawei Noah’s Ark Lab Abstract Large Audio Language Models (LALMs) have made rapid progress on standardized bench- marks, yet their deployment in practical me- dia workflows, curation, archival indexing, and content distribution remains largely unreal- ized. We identify automated audio chapteri- zation, the task of segmenting continuous au- dio streams into thematically coherent chap- ters, as a demanding and commercially conse- quential setting that exposes this gap. Chap- terization is challenging because boundaries are defined less by objective acoustic events than by subjective editorial judgment, requiring models to reason sequentially over long acous- tic contexts and approximate creator-authored boundary decisions. We present AudioChaps, a post-training framework for aligning end-to- end LALMs for this task via Group Relative Policy Optimization (GRPO) guided by Chain- of-Thought (CoT) reasoning. To support train- ing and evaluation, we curate three datasets: AudioChaps-Alignment, derived from creator- annotated chapter boundaries on YouTube; AudioChaps-CoT, which provides structured su- pervision for well-formatted, high-quality, and evidence-grounded boundary reasoning; and AudioChaps-Eval, a held-out benchmark for audio chapterization. Applying GRPO directly without a Supervised Fine-Tuning (SFT) cold start, AudioChaps-R1-Zero already improves average F1 by 33 points over the state-of-the- art LALM Audio-Flamingo-3-Think. The Au- dioChaps framework produces our final aligned LALM, AudioChaps-R1, which improves av- erage F1 by 49 points. These results demon- strate that GRPO-trained LALMs can reliably * Corresponding authors:t.alex@surrey.ac.ukand j.deng16@imperial.ac.uk. † This work was undertaken during Tony Alex’s internship at Huawei Noah’s Ark Lab, London. transform unstructured auditory streams into navigable, structured media. Our code, mod- els, and dataset resources will be released upon acceptance athttps://github.com/ta012/ AudioChaps. 1 Introduction Multimodal Large Language Models (MLLMs), spanning Vision-Language Models (VLMs) (Zhu et al., 2023; Tong et al., 2024), Large Audio Language Models (LALMs) (Deshmukh et al., 2023; Gong et al., 2024; Tang et al., 2024; Alex et al., 2026; Chu et al., 2024; Ghosh et al., 2025), and their variants, have rapidly advanced domain- specific understanding across modalities (Thawkar et al., 2023; Maaz et al., 2023; Brohan et al., 2023). Within the auditory domain in particular, LALMs have evolved through a now-standard recipe of large-scale pretraining, Supervised Fine-Tuning (SFT), and Reinforcement Learning from Human Feedback (RLHF) (Tian et al., 2025), yielding mod- els with increasingly nuanced acoustic comprehen- sion. However, strong performance on benchmark datasets (Sakshi et al., 2024; Ma et al., 2025; Wang et al., 2025) is only a proxy for the broader ambi- tion of intelligence as a utility: the point at which these models are embedded in real-world work- flows and demonstrably improve quality and effi- ciency. Realizing this ambition requires moving beyond generalized capability evaluations toward targeted, task-specific deployments grounded in concrete application domains. The media and entertainment industry offers a particularly compelling testbed for this transfor- mation. The sector has a persistent and grow- ing need for AI-assisted tools, both in the pro- 1 arXiv:2608.16539v1 [cs.SD] 17 Aug 2026 0:00/3:530:00/3:53 ▶ •||||||| |• AudioChaps-R1-8B Chapter 1 Chapter 2 Chapter 3 Chapter N 60s Figure 1: Illustration of the audio chapterization task: the media’s audio is analysed in 60-s chunks via AudioChaps-R1-8B to determine the chapter boundaries, presented to the user as a timeline above the scrubber bar. duction of new content and in the repurposing of archival catalogues for modern distribution plat- forms. Concretely, this includes editorial assis- tance (cut suggestion, subtitling, shot generation), structured metadata creation for indexing and re- trieval, and chapterization for navigable consump- tion. The scale and economic stakes are substantial: legacy broadcasters such as CNN and the BBC must curate decades of archival footage for stream- ing re-release, while platforms including Netflix, YouTube, and Spotify rely on accurate structured metadata to make heterogeneous content libraries efficiently discoverable. Among these tasks, audio chapterization stands out as both broadly applicable and methodologi- cally demanding: it requires reasoning over contin- uous acoustic streams to recover the latent thematic structure that human editors would impose. Given an audio stream of arbitrary duration, the system must partition it into thematically coherent chapters. While narrow speech-only settings such as podcasts can be partially addressed by cascading Automatic Speech Recognition (ASR) with a text-LLM, such pipelines degrade sharply on the heterogeneous audio that defines real-world media: dynamic edi- torial content, music, gaming streams, and beyond. We therefore target an end-to-end LALM that oper- ates directly on raw audio and generalizes across these regimes. The importance of this structuring extends well beyond efficiency. Web search transformed an unstructured collection of pages into a navigable medium and, in doing so, became the substrate on which much of the modern web economy was built. Fine-grained media navigation promises a comparable shift for audio and video: once long- form content is segmented into semantically co- herent units, a further layer of services becomes possible, from precise retrieval and recommenda- tion to personalised education and training, with downstream implications for business, technical upskilling, wellbeing, and healthcare. Chapteriza- tion is a foundational step toward this navigable layer, turning continuous streams into addressable structure on which higher-level applications can build on. A central difficulty is that chapter boundaries are defined less by objective acoustic events than by subjective editorial judgment. Models must therefore reason over distributed acoustic evidence and learn a decision rule that reflects how cre- ators actually structure content. Because uniformly imitating synthetic reasoning traces may pro- duce plausible explanations without yielding well- calibrated boundary decisions, we use GRPO (Shao et al., 2024; Guo et al., 2025) to optimize the model’s final decisions against creator-authored chapter annotations. In our primary instantiation, AudioChaps-R1 is trained in two stages: SFT on the AudioChaps-CoT corpus establishes a struc- tured, evidence-grounded reasoning prior, after which GRPO calibrates boundary decisions toward creator-authored annotations. Our primary contributions are as follows: •We propose AudioChaps, a post-training framework for aligning LALMs for audio chapterization. It uses GRPO to optimize boundary decisions against creator-authored editorial annotations, and the resulting aligned LALM is termed AudioChaps-R1. • We introduce AudioChaps-Alignment, a chapterization dataset derived from creator- annotated YouTube content and stratified across four acoustic regimes (structured speech, dynamic media, gaming, and mu- 2 sic), moving the task beyond the speech- only setting to which existing transcript-based pipelines are limited. •We present AudioChaps-CoT, a reasoning corpus constructed through a novel audio-to- text modality bridge. It provides structured supervision for well-formatted, high-quality, and evidence-grounded boundary reasoning, addressing the absence of supervised reason- ing data for subjective audio chapterization. •We release AudioChaps-Eval, the first bench- mark dedicated to audio-only chapterization, enabling principled comparison on a task with no prior standardized evaluation. •Through extensive experiments, AudioChaps- R1 establishes state-of-the-art performance across all four acoustic regimes, raising aver- age F1 from 28.6 to 77.8 over its base model and surpassing a 32B RL-trained LALM at roughly a quarter of the parameters. 2 Related Works Audio and Video Chapterization. Chapteriza- tion has been studied almost entirely in the video domain. VidChapters-7M (Yang et al., 2023) intro- duced a large corpus of user-chaptered YouTube videos and the tasks of chapter generation, genera- tion given ground-truth boundaries, and grounding. More recent systems extend to hour-long content with richer supervision, combining ASR transcripts and visual tokens under instruction tuning (Pu et al., 2025). Deployed tools likewise remain transcript-driven, running a text-LLM over ASR output, which confines them to speech-dominant material. To the best of our knowledge, no prior work targets chapterization from audio alone. We close this gap with an end-to-end model that oper- ates on raw audio, exploiting non-speech cues such as music and ambient transitions, and that treats boundary placement as editor-supervised acous- tic reasoning rather than transcript segmentation, allowing it to generalize across speech, music, gam- ing, and dynamic media. 3 Methodology In this section, we detail the development and train- ing of AudioChaps using AF3-Think-8B (Goel et al., 2025), a strong contemporary LALM, as our primary backbone. We begin by formally defin- ing the audio chapterization task (Section 3.1). Next, we evaluate the zero-shot chapterization per- formance of AF3-Think-8B to establish the base- line for measuring the improvements introduced by AudioChaps (Section 3.2). We then investigate whether GRPO can directly align the unmodified AF3-Think-8B model with creator-authored chap- ter annotations (Section 3.3). Finally, we present our primary two-stage AudioChaps instantiation, which combines a CoT Supervised Fine-Tuning (SFT) cold start with GRPO-based calibration (Sec- tion 3.4). 3.1 Task Formulation We formulate audio chapterization as boundary de- tection over short audio segments. For training and evaluation, we sample positive and negative 60-second clips from each source video. For pos- itive clips, the ground-truth chapter boundary is sampled uniformly within the central 20 seconds of the window, between 20 and 40 seconds, ensuring at least 20 seconds of acoustic context before and after the transition. Negative clips lie strictly within a single chapter and are sampled with a temporal buffer from any annotated boundary. For each clip, the model is tasked with determining whether a thematic transition is present. This construction provides sufficient pre- and post-transition context for the model to reason over narrative flow rather than rely on instantaneous acoustic cues. At de- ployment, the same model can be applied using a sliding window over audio streams of arbitrary duration (Figure 1; Supplementary Section F), en- abling chapterization without requiring native long- context audio modelling, which remains an open challenge for current LALMs. Ground-truth boundaries are sourced from creator-annotated chapter markers on YouTube, which serve as a naturally occurring proxy for human editorial judgment (Section 4.1). A lim- itation of AF3-Think-8B is that it does not emit boundary timestamps with sufficient accuracy to support reliable temporal reasoning supervision (Section 3.4.1). We therefore formulate the pri- mary training objective as a flow-based, presence- or-absence judgment, keeping both the reasoning traces and final decisions grounded in signals that can be supervised reliably. 3.2 How Much Can AudioChaps Improve an Existing LALM? We use Audio-Flamingo-3-Think-8B (AF3-Think- 8B) (Goel et al., 2025) as the primary backbone 3 throughout our main experiments and first evaluate its zero-shot chapterization performance to estab- lish the unaligned starting point. AF3-Think is the on-demand reasoning variant of Audio Flamingo 3 and does not enforce an explicit reasoning for- mat. Its zero-shot performance therefore provides the principal reference point for quantifying the improvements introduced by AudioChaps. We additionally evaluate Step-Audio-R1- 32B (Tian et al., 2025), a substantially larger audio-reasoning model used to generate the acoustic perception logs during AudioChaps-CoT curation. Its inclusion enables us to assess whether the final AudioChaps-R1 model reflects only the capabilities of a model involved in its supervision pipeline or, through task-specific alignment, develops stronger chapterization ability. Broader comparisons with general-purpose multimodal and omni-modal models are provided in Supplementary Section D. 3.3 The Challenge of Direct RL: AudioChaps-R1-Zero Having established the zero-shot chapterization per- formance of AF3-Think-8B, we next ask whether reinforcement learning alone can align its decisions with the editorial judgment reflected in creator- authored annotations. Recent work on text-only reasoning models has shown that GRPO-style optimization with rule-based rewards can elicit substantial reasoning capabilities from unaligned base models, an approach popularized as “R1- Zero” (Guo et al., 2025). We investigate the analo- gous setting for audio by applying GRPO directly to AF3-Think-8B, yielding AudioChaps-R1-Zero. A direct port of the R1-Zero recipe is, how- ever, impractical in our setting. AF3-Think-8B does not natively emit responses in the strict <think>...</think><answer>...</answer> schema; enforcing this format via reward would penalize nearly all initial rollouts, collapsing the learning signal. We therefore adapt the reward function to the model’s native output style: rather than requiring exact tag-based formatting, we reward (i) the presence of a discernible reasoning trace preceding the final decision and (i) the correctness of the binary boundary verdict. This formulation preserves the spirit of rule-based RL while remaining tractable for an unaligned starting point. While direct GRPO yields measurable gains over the AF3-Think-8B baseline, qualitative in- spection reveals inconsistent reasoning styles and weak grounding in task-relevant acoustic evidence across regimes. For our primary AF3-based instan- tiation, we therefore introduce a CoT cold start be- fore GRPO. This stage standardizes the output for- mat and provides an evidence-grounded reasoning initialization, after which GRPO calibrates bound- ary decisions against creator-authored annotations. 3.4 The AudioChaps Framework This section presents the primary instantiation of AudioChaps on AF3-Think-8B. We first describe the construction of AudioChaps-CoT through an audio-to-text modality bridge (Section 3.4.1), and then detail the SFT cold start and GRPO calibration used to obtain AudioChaps-R1 (Section 3.4.2). 3.4.1 CoT Data Generation for Audio Chapterization Directly applying GRPO requires the model to discover both the desired reasoning strategy and output format from sparse binary rewards. To provide a structured initialization, we construct AudioChaps-CoT through an audio-to-text modal- ity bridge (Figure 2). The pipeline converts contin- uous audio into a sanitized acoustic perception log that a stronger text-based reasoning model can use to generate evidence-grounded CoT targets. Stage 1: Pseudo-CoT Generation. For each clip, Step-Audio-R1 generates an initial pseudo- CoT conditioned on the audio, subtype, video title, and creator-authored chapter supervision. Positive samples are grounded using the semantic change between adjacent chapter titles, while negative sam- ples are identified as continuous segments. This supervision encourages the model to identify task- relevant acoustic evidence rather than produce a generic audio description. Stage 2: Sanitized Acoustic Perception Log. Because the pseudo-CoT contains explicit label and metadata supervision, it cannot be used directly for SFT. Step-Audio-R1 therefore processes the au- dio again, using the pseudo-CoT only as contextual guidance, and produces a chronological description of audible evidence. Explicit verdicts and structural terms such as “boundary,” “transition,” “segment,” and “chapter” are removed, yielding a perception log that preserves relevant acoustic cues without revealing the target label. Stage 3: Final CoT Synthesis. Gemini 2.5 Pro (Comanici et al., 2025) receives only the sani- 4 Stage 1: Boundary-Relevant Pseudo-CoT Generation Inputs • Raw audio clip • Creator-authored chapter labels • Video metadata • Subtype-specific hints Model: Step-Audio-R1 Output: Boundary-relevant pseudo-CoT Note: May contain label-revealing information Stage 2: Pseudo-CoT Sanitization Inputs • Raw audio clip • Subtype-specific hints • Boundary-relevant pseudo-CoT from Stage 1 Model: Step-Audio-R1 Output: Sanitized acoustic perception log Note: Removes label-revealing and structural language Stage 3: Final CoT Synthesis Inputs • Sanitized acoustic perception log from Stage 2 • Chapter-boundary query Model: Gemini 2.5 Pro Output: Structured CoT target <think>...</think> <answer>...</answer> Note: Generates the final AudioChaps-CoT Figure 2: AudioChaps-CoT curation pipeline. A boundary-relevant pseudo-CoT is generated, sanitized into an acoustic perception log, and then converted into the final structured CoT target used for the SFT cold start. tized perception log and the binary chapterization query. It infers whether a boundary is present and generates the final target in the required <think>...</think><answer>...</answer> format.The synthetic CoT targets describe the evolution of the acoustic evidence using flow-based language rather than exact timestamps. This design is particularly suitable for LALMs such as AF3-Think-8B, which we found did not predict temporal locations consistently enough to support reliable timestamp-based supervision. The resulting corpus provides well-formatted, evidence-grounded targets for the SFT cold start. 3.4.2 SFT Cold Start and GRPO Calibration Equipped with the AudioChaps-CoT corpus, we align AF3-Think-8B through a two-stage proce- dure: an SFT cold start followed by GRPO-based calibration. Stage 1: Cold-Start SFT. AF3-Think-8B does not reliably produce well-formatted, high-quality reasoning traces for boundary detection. We there- fore fine-tune it on AudioChaps-CoT (prompt tem- plate in Supplementary Section A.1) to obtain AudioChaps-SFT. This stage establishes the struc- tured output format required for rule-based GRPO rewards, consistently separating the reasoning trace from the final verdict, while providing a coherent initialization for subsequent calibration. Stage 2: GRPO-Based Calibration. Although SFT provides a structured initialization, it encour- ages the model to imitate a single synthetic ratio- nale for each example. This can produce plausible, well-formatted explanations without adequately calibrating when a boundary should be predicted. We therefore apply GRPO to AudioChaps-SFT, sampling a group ofGrollouts for each query, where each rollout may follow a different reason- ing trajectory before producing a boundary ver- dict. GRPO assigns higher relative advantage to tra- jectories whose final decisions agree with creator- authored chapter annotations, thereby calibrating the model’s boundary judgments. Unlike SFT, which learns from one prescribed rationale, GRPO can reinforce multiple reasoning trajectories that lead to the same creator-aligned decision. This flexibility is well suited to the subjective nature of chapterization, where the same editorial bound- ary may be supported by different combinations of semantic, discourse, acoustic, and structural cues. Formally, as shown in Equation 1, we optimize the policyπ θ over trajectorieso i G i=1 sampled conditional on the audio queryq. Here,π θ old de- notes the policy used to generate the current rollout group, whileπ ref is the frozen AudioChaps-SFT reference policy. The clipped surrogate objective bounds the policy ratio within[1−ε, 1+ε], and the Kullback–Leibler penaltyβD KL (π θ ∥ π ref )limits divergence from the SFT initialization. For each rollouto i , we compute an unweighted scalar rule-based reward: r i = r format (o i ) + r accuracy (o i ),(2) wherer format verifies the presence and validity of the<think>and<answer>tags, andr accuracy mea- sures agreement between the predicted boundary decision and the creator-authored label. Both com- ponents are binary, taking value1when the corre- sponding criterion is satisfied and 0 otherwise. The group-relative advantage is obtained by nor- malizing each reward within theGrollouts sampled for the same query: A i = r i − μ r σ r + δ , μ r = 1 G G X j=1 r j ,(3) whereσ r is the standard deviation of the group rewards andδis a small constant for numerical stability. A rollout therefore receives positive ad- vantage when its reward exceeds that of the other reasoning trajectories sampled for the same audio query. This rule-based formulation enforces valid structured outputs while directly aligning bound- ary decisions with editorial supervision, without requiring a learned reward model. 5 J GRPO (θ) =E q∼P(Q),o i G i=1 ∼π θ old (O|q) " 1 G G X i=1 min π θ (o i | q) π θ old (o i | q) A i , clip π θ (o i | q) π θ old (o i | q) , 1− ε, 1 + ε A i − βD KL (π θ ∥ π ref ) # . (1) ModelscAccPreRecF1 DM60.255.362.758.8 Step-Audio-R1-32BG60.159.153.456.1 (Zero-Shot)M62.453.151.952.5 S65.670.169.769.9 Avg62.159.459.459.3 DM55.847.523.731.6 AF3-Think-8BG50.145.954.649.9 (Zero-Shot)M62.364.03.26.0 S44.746.419.127.0 Avg53.251.025.228.6 DM75.970.077.273.4 AudioChaps-R1-8BG77.675.175.975.5 (Ours)M88.183.685.684.6 S73.671.086.077.8 Avg78.874.981.277.8 Table 1: Performance comparison on AudioChaps-Eval across the four audio chapterization subcategories: Dy- namic Media (DM), Gaming (G), Music (M), and Struc- tured Speech (S). Accuracy (Acc), precision (Pre), re- call (Rec), and F1 are reported as percentages, together with the macro-average (Avg) over subcategories. De- tailed comparison with additional multimodal and omni- modal LLMs is provided in Supplementary Table 7. Overall, SFT supplies the structured reasoning interface required for stable optimization, while GRPO provides the decisive alignment signal that calibrates boundary judgments to creator-authored editorial supervision. 4 Experiments 4.1 Dataset Data Source and Acoustic Stratification. We curate our training and evaluation data from VidChapters-7M (Yang et al., 2023), a large-scale collection of YouTube videos with chapter bound- aries manually provided by content creators. To evaluate audio chapterization across heterogeneous media, we query the YouTube Data API for each video’s category label and group the selected con- tent into four broad acoustic subtypes: structured speech (Education, Science & Technology, and Howto & Style), dynamic media (Entertainment, Comedy, and Film & Animation), gaming, and mu- sic. This stratification exposes the model to diverse acoustic conditions, ranging from narrated instruc- tional content to edited entertainment, gameplay audio, and music-dominant recordings. Clip Construction. For each source video, we construct positive and negative 60-second clips. A positive clip contains a creator-authored chapter boundary, with the boundary sampled uniformly between 20 and 40 seconds of the clip. This place- ment provides at least 20 seconds of acoustic con- text before and after the annotated transition. Neg- ative clips are sampled entirely within a single an- notated chapter and are kept temporally separated from annotated boundaries. Where possible, nega- tive clips are sampled from the same source videos as positive clips, reducing differences in speakers, recording conditions, and production style between the two classes. Training Data. The resulting AudioChaps- Alignment training set contains approximately 30k labelled clips across the four acoustic subtypes, in- cluding 13,347 positive boundary-spanning clips and 16,636 negative within-chapter clips. A 22k- clip subset is used to construct AudioChaps-CoT for the SFT cold start, while the labelled align- ment data provides creator-authored supervision for GRPO. Clip-Level Evaluation. For evaluation, we con- struct AudioChaps-Eval, a held-out test set con- taining approximately 16k clips from 749 source videos, including 7,011 positive clips and 8,952 negative clips. The training and test partitions are split at the source-video level, so no source video appears in both sets. Detailed per-subtype counts and class distributions are provided in Supplemen- tary Section G. Full-Length Evaluation. In addition to the con- trolled clip-level AudioChaps-Eval benchmark, we evaluate chapter-boundary detection on a balanced set of full-length recordings spanning all four acoustic subtypes. The set comprises 40 record- ings, with 10 from each subtype, and contains 387 reference boundaries. The recordings have a mean duration of 49 minutes, yielding approxi- mately 33 hours of long-form audio in total. The continuous-stream inference and evaluation proto- 6 col is described in Section 4.2 and Supplementary Section F. 4.2 Results Discussion Table 1 compares two zero-shot baselines with our final aligned model on AudioChaps-Eval. AF3- Think-8B (Goel et al., 2025) serves as the back- bone for AudioChaps-R1, while Step-Audio-R1- 32B (Tian et al., 2025) is used to generate the acoustic perception logs during AudioChaps-CoT curation. Importantly, the boundary labels are pro- vided by neither model and remain grounded in creator-authored annotations. Comparing against both models therefore evaluates the gains over the underlying backbone as well as over the larger model involved in the supervision pipeline. Ad- ditional comparisons with multimodal and omni- modal LLMs are provided in Supplementary Ta- ble 7. Zero-Shot LALMs Find Chapterization Diffi- cult. AF3-Think-8B exhibits a conservative pre- diction pattern: on Music, it achieves 64.0 preci- sion but only 3.2 recall, indicating that it predomi- nantly predicts “no boundary” and identifies only the clearest transitions. The same tendency appears in Structured Speech and Dynamic Media, where recall remains below 25. Step-Audio-R1-32B pro- duces a more balanced precision–recall profile, but its accuracy remains between 60% and 65% across all four subtypes, indicating that zero-shot chapteri- zation remains challenging even for a substantially larger RL-trained LALM. Comparison with an ASR–LLM Cascade. To test whether chapter boundaries can be identified from transcripts alone, we construct a cascade base- line using Whisper-Large-V3 (Radford et al., 2022) for automatic speech recognition(ASR), followed by Qwen3-235B-A22B-Instruct-2507-FP8 (Yang et al., 2025) to analyse the transcript and determine whether a chapter boundary is present. We evalu- ate this baseline on the Structured Speech subset, where ASR is expected to be most reliable. The cascade achieves 71.4 precision, 36.1 recall, and 48.0 F1, compared with 71.0 precision, 86.0 recall, and 77.8 F1 for AudioChaps-R1. The comparable precision but substantially lower recall suggests that transcripts capture some explicit semantic tran- sitions but miss boundaries signalled by acoustic pacing, delivery changes, pauses, and background continuity. AudioChaps-R1 outperforms both its backbone LALM and the larger LALM used for CoT cura- tion.AudioChaps-R1 yields substantial and con- sistent gains over the AF3-Think-8B backbone. Av- eraged across subtypes, accuracy rises from 53.2 to 78.8, precision from 51.0 to 74.9, recall from 25.2 to 81.2, and F1 from 28.6 to 77.8. A paired video- level bootstrap analysis with 10,000 resamples con- firms that AudioChaps-R1 improves macro-F1 over AF3-Think-8B by 49.2 points, with a 95% confi- dence interval of [46.2, 52.2] and a two-sided boot- strap p-value below10 −4 . Most of the improve- ment comes from recall, reflecting the backbone’s conservative tendency to predict no boundary. The largest gain occurs on Music, where recall rises from 3.2 to 85.6 and F1 from 6.0 to 84.6. These re- call gains do not come at the expense of precision, which also increases across all subtypes, includ- ing from 47.5 to 70.0 on Dynamic Media and from 46.4 to 71.0 on Structured Speech. The largest accu- racy gain occurs on Structured Speech, increasing from 44.7 to 73.6. AudioChaps-R1 also outper- forms Step-Audio-R1-32B, the larger LALM used to produce the acoustic perception logs during CoT curation, by 14.6, 19.4, 32.1, and 7.9 F1 points on Dynamic Media, Gaming, Music, and Structured Speech, respectively, despite using approximately one quarter of its parameters. Its improvement over Step-Audio-R1-32B indicates that AudioChaps-R1 does not merely inherit the capabilities of a model involved in its supervision pipeline, but develops stronger chapterization ability through task-specific alignment. AudioChaps-R1 performs consistently across acoustic regimes. AF3-Think-8B exhibits sub- stantial variation across subtypes, spanning 43.9 F1 points from 6.0 on Music to 49.9 on Gaming. In contrast, AudioChaps-R1 narrows this range to 11.2 points, from 73.4 on Dynamic Media to 84.6 on Music. This more consistent performance in- dicates that the aligned model remains effective across diverse acoustic regimes rather than favour- ing a single content type. Full-length chapter-boundary detection. For full-length evaluation, we apply the model using a sliding 60-second window with a 20-second hop (Figure 1). Consecutive windows predicted as con- taining a boundary are grouped into a positive run, from which a single boundary estimate is placed at the temporal midpoint. As shown in Table 2, the zero-shot AF3-Think- 7 ModelF1PreRecdev-R2E Fixed interval, 180 s9.59.011.649.0 s AF3-Think-8B6.58.69.638.0 s AudioChaps-R1-8B37.634.145.710.0 s Table 2: Full-length chapter-boundary detection on un- cropped audio. Predictions are matched one-to-one with reference boundaries under a±10-second tolerance. Precision, recall, and F1 are computed per recording, macro-averaged within each acoustic subtype, and then averaged equally across the four subtypes.dev-R2E denotes the pooled median temporal distance from each reference boundary to its nearest predicted boundary. Detailed results are provided in Supplementary Table 8. ModelDMGMSS AF3-Think-8B31.649.96.027.0 AudioChaps-R1-Zero-8B57.864.459.162.9 AudioChaps-SFT-8B70.072.574.177.9 AudioChaps-R1-8B73.475.584.677.8 Table 3: Ablation of the AudioChaps training stages. F1 scores are reported for Dynamic Media (DM), Gaming (G), Music (M), and Structured Speech (S). Detailed results are provided in Supplementary Section B, Ta- ble 5. 8B backbone achieves an F1 score of 6.5 on un- cropped, full-length recordings. A fixed-interval baseline that places a boundary every 180 seconds reaches 9.5 F1, whereas AudioChaps-R1-8B in- creases F1 to 37.6 and reduces the pooled median reference-to-estimate deviation from 38.0 to 10.0 seconds. These results show that local boundary decisions can be consolidated into effective recording-level chapter predictions, providing a practical approach to long-form audio chapterization without relying on native long-context audio modelling, which re- mains limited in current LALMs. Additional de- tails are provided in Supplementary Section F. 4.3 Ablation Study Table 3 isolates the contribution of each training stage by comparing the AF3-Think-8B backbone, direct GRPO without an SFT cold start, SFT alone, and the full SFT+GRPO pipeline. F1 is reported across Dynamic Media, Gaming, Music, and Struc- tured Speech, with full per-metric results provided in Supplementary Section B, Table 5. All variants use AF3-Think-8B as the underlying backbone. Direct RL yields substantial but uneven gains. Applying GRPO directly to the AF3-Think-8B backbone produces AudioChaps-R1-Zero-8B, rais- ing F1 to 57.8, 64.4, 59.1, and 62.9 on Dynamic Media, Gaming, Music, and Structured Speech, re- spectively. These correspond to gains of 26.2, 14.5, 53.1, and 35.9 points over the backbone. Much of the improvement is driven by increased recall, in- dicating that direct GRPO reduces the backbone’s tendency to default to “no boundary.” Notably, this recall-oriented shift resembles the trend produced by SFT, but remains consistently weaker across all four subtypes. This suggests that, without a struc- tured reasoning prior, GRPO exploration primarily shifts the model toward greater boundary sensitivity rather than producing a sufficiently well-calibrated chapterization policy. These results support the use of a CoT cold start before applying GRPO to AF3-Think-8B. SFT cold-start establishes a high-recall prior. AudioChaps-SFT-8B improves over AudioChaps- R1-Zero across all four subtypes, reaching F1 scores of 70.0, 72.5, 74.1, and 77.9 on Dynamic Media, Gaming, Music, and Structured Speech, re- spectively. The improvement is primarily recall- driven, with recall rising to 88.5, 83.8, 90.1, and 90.3, while precision remains comparatively lower. On Music, for example, the model achieves 90.1 re- call but 62.9 precision. This indicates that imitating the AudioChaps-CoT targets establishes a strong high-recall prior, providing an effective starting point for subsequent GRPO calibration. GRPO calibration delivers the best overall trade-off.Comparing AudioChaps-SFT-8B with AudioChaps-R1-8B isolates the role of GRPO once a structured reasoning prior has been established through the CoT cold start. Rather than simply increasing boundary sensitivity, GRPO corrects the permissive operating point induced by SFT: preci- sion rises by 12.1 points on Dynamic Media, 11.2 on Gaming, 20.7 on Music, and 2.5 on Structured Speech, while recall decreases to varying degrees. The benefit is largest where SFT leaves the most precision headroom. On Music, GRPO trades 4.5 points of recall for a 20.7-point precision gain, in- creasing F1 by 10.5 points. On Structured Speech, SFT is already well calibrated, so the 2.5-point pre- cision gain is offset by a 4.2-point recall reduction and F1 remains effectively unchanged. Overall, AudioChaps-R1-8B achieves the best F1 on three of the four subtypes and remains effectively tied with AudioChaps-SFT-8B on Structured Speech. These results clarify how GRPO operates once 8 an evidence-grounded reasoning initialization and a consistent output format have been established. SFT teaches the model to interpret the audio in relation to the chapter-boundary question and to produce well-structured reasoning, while GRPO improves precision by calibrating the resulting de- cisions toward creator-authored annotations, pri- marily by suppressing unsupported boundary calls. A similar precision-oriented shift is observed with the MOSS-Think backbone, as discussed below. Full results are provided in Supplementary Table 5. Backbone generalization of AudioChaps. Au- dioChaps is designed not as a replacement for a particular base LALM, but as a backbone-agnostic GRPO-based alignment framework that strength- ens the chapterization ability already present in that model.To evaluate whether its gains extend beyond AF3-Think-8B, we apply Au- dioChaps to MOSS-Audio-8B-Thinking (MOSS- Think-8B) (Yang et al., 2026), which exhibits strong zero-shot chapterization performance. Be- cause MOSS-Think-8B has already undergone su- pervised instruction tuning, an explicit reasoning cold start, and reinforcement learning for audio rea- soning, it already provides a structured reasoning prior and consistent reasoning format. We therefore apply the AudioChaps GRPO alignment stage di- rectly. AudioChaps further improves MOSS-Think- 8B across all four acoustic subtypes, increasing F1 from 69.3 to 79.2 on Dynamic Media, 72.9 to 78.2 on Gaming, 75.7 to 85.6 on Music, and 76.5 to 82.8 on Structured Speech. The gain is accompa- nied by the same precision-oriented shift observed in the AF3-Think-8B stage-wise ablation: average precision rises from 63.3 to 82.4, while recall de- creases from 88.0 to 80.6, increasing average F1 from 73.6 to 81.4. These results show that, when a suitable reasoning prior and output format are already present, task-specific GRPO can further im- prove the native chapterization ability of a capable LALM by suppressing unsupported boundary calls and calibrating predictions toward creator-authored annotations. Full results are provided in Table 6 and Section C of the supplementary material. Comparisonwithadditionalmultimodal and omni LLMs. We further compare Au- dioChaps against two strong zero-shot baselines on AudioChaps-Eval: Qwen3-Omni-30B-A3B- Thinking (Xu et al., 2025) and Gemini 2.5 Flash (Comanici et al., 2025).MOSS-Think- AudioChaps-R1-8B outperforms both larger general-purpose models, achieving an average F1 of 81.44, compared with 76.75 for Gemini 2.5 Flash and 75.29 for Qwen3-Omni.This corresponds to gains of 4.69 and 6.15 F1 points, respectively. These results show that AudioChaps enables an 8B-parameter LALM to outper- form larger zero-shot models by aligning its chapterization decisions with creator-authored annotations. 5 Conclusion We study audio chapterization as a concrete set- ting in which LALMs can move beyond bench- mark performance toward practical media work- flows. By treating chapter-boundary placement as editorial judgment rather than a fixed acoustic event, we align end-to-end LALMs with creator- authored chapter annotations using three resources: AudioChaps-Alignment, AudioChaps-CoT, and AudioChaps-Eval. Our two-stage training recipe combines a CoT SFT cold start, which establishes a structured high-recall reasoning prior, with GRPO, which improves precision by suppressing unsup- ported boundary calls and calibrating predictions toward creator-authored annotations. AudioChaps- R1 substantially improves over its AF3-Think-8B backbone and surpasses the larger Step-Audio-R1 model used to produce its acoustic perception su- pervision, despite using roughly one quarter of the parameters. Limitations.The current framework does not na- tively support long-context audio modelling. This limits its ability to generate globally informed chap- ter titles and segment summaries. Extending Au- dioChaps to native long-context modelling and full chapter generation is therefore an important direc- tion toward deployment in production media work- flows. Acknowledgments This research was supported by the EPSRC and BBC Prosperity Partnership “AI4ME: Future Per- sonalized Object Based Media Experiences Deliv- ered at Scale Anywhere” (EP/V038087/1). The authors also acknowledge the computational re- sources provided by Isambard-AI, part of the Na- tional AI Research Resource (AIRR) (McIntosh- Smith et al., 2024). Isambard-AI is operated by the University of Bristol and funded by the UK Gov- ernment’s Department for Science, Innovation and 9 Technology (DSIT) through UK Research and In- novation (UKRI) and the Science and Technology Facilities Council (STFC), under grant ST/AIRR/I- A-I/1023. References Tony Alex, Wish Suharitdamrong, Sara Atito, Armin Mustafa, Philip J. B. Jackson, Imran Razzak, and Muhammad Awais. 2026. Pal: Probing audio en- coders via llms – audio information transfer into llms. Preprint, arXiv:2506.10423. Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Xi Chen, Krzysztof Choromanski, Tianli Ding, Danny Driess, Avinava Dubey, Chelsea Finn, and 1 others. 2023. Rt-2: Vision-language-action models transfer web knowledge to robotic control. arXiv preprint arXiv:2307.15818. Yunfei Chu, Jin Xu, Qian Yang, Haojie Wei, Xipin Wei, Zhifang Guo, Yichong Leng, Yuanjun Lv, Jinzheng He, Junyang Lin, Chang Zhou, and Jingren Zhou. 2024.Qwen2-audio technical report. Preprint, arXiv:2407.10759. Gheorghe Comanici and 1 others. 2025. Gemini 2.5: Pushing the frontier with advanced reasoning, multi- modality, long context, and next generation agentic capabilities. Preprint, arXiv:2507.06261. Soham Deshmukh, Benjamin Elizalde, Rita Singh, and Huaming Wang. 2023. Pengi: An audio language model for audio tasks. Advances in Neural Informa- tion Processing Systems, 36. Sreyan Ghosh, Zhifeng Kong, Sonal Kumar, S Sak- shi, Jaehyeon Kim, Wei Ping, Rafael Valle, Di- nesh Manocha, and Bryan Catanzaro. 2025. Audio flamingo 2: An audio-language model with long- audio understanding and expert reasoning abilities. arXiv preprint arXiv:2503.03983. Arushi Goel, Sreyan Ghosh, Jaehyeon Kim, Sonal Ku- mar, Zhifeng Kong, Sang-gil Lee, Chao-Han Huck Yang, Ramani Duraiswami, Dinesh Manocha, Rafael Valle, and 1 others. 2025. Audio flamingo 3: Advanc- ing audio intelligence with fully open large audio language models. arXiv preprint arXiv:2507.08128. Yuan Gong, Hongyin Luo, Alexander H. Liu, Leonid Karlinsky, and James R. Glass. 2024. Listen, think, and understand. In The Twelfth International Confer- ence on Learning Representations. Daya Guo, Yang, and 1 others. 2025. Deepseek-r1 incentivizes reasoning in llms through reinforcement learning. Nature, 645(8081):633–638. Ziyang Ma, Yinghao Ma, Yanqiao Zhu, Chen Yang, Yi-Wen Chao, Ruiyang Xu, Wenxi Chen, Yuanzhe Chen, Zhuo Chen, Jian Cong, and 1 others. 2025. Mmar: A challenging benchmark for deep reasoning in speech, audio, music, and their mix. arXiv preprint arXiv:2505.13032. Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. 2023. Video-chatgpt: Towards detailed video understanding via large vision and language models. arXiv preprint arXiv:2306.05424. Simon McIntosh-Smith, Sadaf R Alam, and Christo- pher Woods. 2024. Isambard-ai: a leadership class supercomputer optimised specifically for artificial intelligence. Preprint, arXiv:2410.11199. Junfu Pu, Teng Wang, Yixiao Ge, Yuying Ge, Chen Li, and Ying Shan. 2025. Arc-chapter: Structuring hour- long videos into navigable chapters and hierarchical summaries. Preprint, arXiv:2511.14349. Alec Radford, Jong Wook Kim, Tao Xu, Greg Brock- man, Christine McLeavey, and Ilya Sutskever. 2022. Robust speech recognition via large-scale weak su- pervision. Preprint, arXiv:2212.04356. S Sakshi, Utkarsh Tyagi, Sonal Kumar, Ashish Seth, Ramaneswaran Selvakumar, Oriol Nieto, Ramani Duraiswami, Sreyan Ghosh, and Dinesh Manocha. 2024. Mmau: A massive multi-task audio under- standing and reasoning benchmark. arXiv preprint arXiv:2410.19168. Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, Y. K. Li, Y. Wu, and Daya Guo. 2024. Deepseekmath: Pushing the limits of mathemati- cal reasoning in open language models. Preprint, arXiv:2402.03300. Changli Tang, Wenyi Yu, Guangzhi Sun, Xianzhao Chen, Tian Tan, Wei Li, Lu Lu, Zejun MA, and Chao Zhang. 2024. SALMONN: Towards generic hearing abilities for large language models. In The Twelfth International Conference on Learning Representa- tions. Omkar Thawkar, Abdelrahman Shaker, Sahal Shaji Mul- lappilly, Hisham Cholakkal, Rao Muhammad Anwer, Salman Khan, Jorma Laaksonen, and Fahad Shahbaz Khan. 2023. Xraygpt: Chest radiographs summariza- tion using medical vision-language models. arXiv preprint arXiv:2306.07971. Fei Tian, Xiangyu Tony Zhang, Yuxin Zhang, Haoyang Zhang, Yuxin Li, Daijiao Liu, Yayue Deng, Dong- hang Wu, Jun Chen, Liang Zhao, Chengyuan Yao, Hexin Liu, Eng Siong Chng, Xuerui Yang, Xiangyu Zhang, Daxin Jiang, and Gang Yu. 2025. Step-audio- r1 technical report. Preprint, arXiv:2511.15848. Peter Tong, Ellis Brown, Penghao Wu, Sanghyun Woo, Adithya Jairam Vedagiri IYER, Sai Charitha Akula, Shusheng Yang, Jihan Yang, Manoj Middepogu, Ziteng Wang, and 1 others. 2024. Cambrian-1: A fully open, vision-centric exploration of multimodal llms. Advances in Neural Information Processing Systems, 37:87310–87356. Dingdong Wang, Jincenzi Wu, Junan Li, Dongchao Yang, Xueyuan Chen, Tianhua Zhang, and Helen 10 Meng. 2025. Mmsu: A massive multi-task spoken language understanding and reasoning benchmark. arXiv preprint arXiv:2506.04779. Jin Xu, Zhifang Guo, Hangrui Hu, Yunfei Chu, Xiong Wang, Jinzheng He, Yuxuan Wang, Xian Shi, Ting He, Xinfa Zhu, Yuanjun Lv, Yongqi Wang, Dake Guo, He Wang, Linhan Ma, Pei Zhang, Xinyu Zhang, Hongkun Hao, Zishan Guo, and 19 oth- ers. 2025. Qwen3-omni technical report. Preprint, arXiv:2509.17765. An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, and 1 others. 2025.Qwen3 technical report. arXiv preprint arXiv:2505.09388. Antoine Yang, Arsha Nagrani, Ivan Laptev, Josef Sivic, and Cordelia Schmid. 2023. Vidchapters-7m: Video chapters at scale. Preprint, arXiv:2309.13952. Chen Yang, Chufan Yu, Hanfu Chen, Jie Zhu, Jingqi Chen, Ke Chen, Wenxuan Wang, Yang Wang, Yaozhou Jiang, Yi Jiang, Zhengyuan Lin, Ziqi Chen, Zhaoye Fei, Chenghao Liu, Donghua Yu, Jun Zhan, Kang Yu, Kexin Huang, Liwei Fan, and 11 oth- ers. 2026. Moss-audio technical report. Preprint, arXiv:2606.01802. Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. 2023. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592. HyperparameterSFTGRPO Epochs21 Learning rate1e−61e−6 Per-device batch size11 Gradient accumulation41 Max completion length—768 Generations per prompt (G)—8 KL coefficient (β)—0.04 Max gradient norm55 Precisionbf16bf16 AttentionFA-2FA-2 DeepSpeedZeRO-2ZeRO-2 Table 4: Training hyperparameters for the SFT cold- start and GRPO alignment stages. A Training Details A.1 Training Prompt Template Across both training stages, we adopt a unified prompt template that frames chapterization as a single-choice question. Each prompt presents the binary query introduced in Section 3.1, augmented with two prescriptive elements: (i) an instruction to engage in deliberate, human-like internal re- flection, using natural cognitive markers such as “let me think,” “wait,” or “let’s break it down,” and to perform self-verification before commit- ting to a decision; and (i) a strict output specifica- tion requiring the reasoning trace to be enclosed within<think>...</think>tags and the final an- swer, restricted to a single option letter, within <answer>...</answer>tags. The full prompt is shown in Figure 3. A.2 Hyperparameters and Training Budget Table 4 lists the hyperparameters for both train- ing stages. In terms of compute, the SFT stage was run on a single node with 8 NVIDIA H200- 140GB GPUs and completed in approximately 4 hours. GRPO alignment was run on 8 nodes, each with 4 NVIDIA GH200-96GB, and completed in approximately 10 hours. Both stages fine-tune the 8B-parameter backbone rather than training from scratch, keeping the overall budget modest relative to pretraining a model of comparable capability. B Detailed Ablation Results This section provides the full per-subtype, per- metric breakdown behind the F1 summary reported in the main paper Table 3. Table 5 lists accuracy, precision, recall, and F1 for the AF3-Think-8B base and the three variants of our pipeline (R1- Zero, SFT, and the full R1 model) across all four 11 Training prompt template Is there a chapter boundary present in this audio? Choose one from the following choices: (A) Yes (B) No Please think about this question as if you were a human pondering deeply. Engage in an internal dialogue using expressions such as ‘let me think’, ‘wait’, ‘Hmm’, ‘oh, I see’, ‘let’s break it down’, etc, or other natural language thought expressions. It’s encouraged to include self-reflection or verification in the reasoning process. Provide your detailed reasoning between the <think> </think> tags, and then give your final answer between the <answer> </answer> tags. Please provide only the single option letter (e.g., A, B) within the <answer> </answer> tags. Figure 3: The unified prompt template used in both the SFT cold-start and GRPO stages. The model is asked a binary single-choice question, instructed to reason with natural reflective markers inside<think>tags, and to emit a single option letter inside <answer> tags. content subcategories, allowing the precision and recall behaviour discussed in Section 4.3 to be read off directly. C Backbone Generalization of AudioChaps Our primary contribution is not a specific back- bone, but a GRPO-based framework for adapt- ing LALMs to creator-authored editorial anno- tations for audio chapterization. We instantiate AudioChaps on both AF3-Think-8B (Goel et al., 2025) and MOSS-Audio-8B-Thinking (MOSS- Think-8B) (Yang et al., 2026), adapting the ini- tialization stage to the capabilities of each model. AF3-Think-8B requires a CoT SFT cold start to es- tablish the target reasoning format before GRPO. In contrast, MOSS-Think-8B has already undergone an explicit reasoning cold start and reinforcement learning for audio reasoning. It therefore provides a strong reasoning prior and produces outputs in the required structured format, allowing us to apply the AudioChaps GRPO stage directly. A consistent precision-recall trend emerges across the two backbones. As shown in Table 5, the AF3-based SFT model is strongly recall-oriented, achieving 88.2 macro-average recall but only 63.3 precision. Subsequent GRPO calibration raises precision to 74.9 and F1 from 73.6 to 77.8, while reducing recall to 81.2. The same behaviour is ob- served with MOSS-Think-8B (Table 6): the base LALM (MOSS-Think-8B) achieves 88.0 recall and 63.3 precision, whereas the AudioChaps-adapted model (MOSS-Think-AudioChaps-R1-8B) raises precision to 82.4 and F1 from 73.6 to 81.4, with recall decreasing to 80.6. The MOSS-based AudioChaps model(MOSS- Think-AudioChaps-R1-8B) improves F1 across all four subcategories, with gains of 9.9 points on Dy- namic Media, 5.3 on Gaming, 9.9 on Music, and 6.3 on Structured Speech. These results indicate that GRPO plays a similar calibration role across distinct LALM families: it corrects an overly recall- heavy decision policy and produces a more selec- tive and balanced chapter-boundary detector. At the same time, these results show that AudioChaps can further improve the native chapterization ability of different LALMs, regardless of their initial ca- pabilities, by calibrating their boundary decisions toward creator-authored annotations. D Comparison with Larger General-Purpose Multimodal and Omni-Modal Models We further compare AudioChaps with two more recent and substantially larger audio-language mod- els, Gemini 2.5 Flash (Comanici et al., 2025) and Qwen3-Omni-30B-A3B-Thinking (Xu et al., 2025). Table 7 shows that the 8B-parameter Au- dioChaps models remain competitive despite their smaller scale. AudioChaps-R1 with the MOSS backbone achieves the best macro-average ac- curacy, precision, and F1, reaching 83.4, 82.4, and 81.4, respectively.Compared with Gem- ini 2.5 Flash, it improves macro-average F1 by 4.6 points, and compared with Qwen3-Omni-30B- A3B-Thinking, it improves F1 by 6.1 points. Gem- ini obtains the highest macro-average recall at 82.7, indicating a more recall-oriented prediction policy, whereas MOSS-Think-AudioChaps-R1-8B provides a better precision–recall balance. The advantage of MOSS-Think-AudioChaps-R1-8B is consistent across subcategories, achieving the high- 12 ModelscAccPre RecF1 DM 55.8 47.5 23.7 31.6 AF3-Think-8BG50.1 45.9 54.6 49.9 (Base-LALM)M62.3 64.03.26.0 S44.7 46.4 19.1 27.0 Avg 53.2 51.0 25.2 28.6 DM 61.6 55.0 61.0 57.8 AudioChaps-R1-G65.3 60.3 69.2 64.4 Zero-8BM75.6 82.1 46.1 59.1 S60.0 62.8 62.9 62.9 Avg 65.6 65.1 59.8 61.1 DM 67.3 57.9 88.5 70.0 AudioChaps-SFT-G71.1 63.9 83.8 72.5 8BM75.9 62.9 90.1 74.1 S72.5 68.5 90.3 77.9 Avg 71.7 63.3 88.2 73.6 DM 75.9 70.0 77.2 73.4 AudioChaps-R1-G77.6 75.1 75.9 75.5 8BM88.1 83.6 85.6 84.6 S73.6 71.0 86.0 77.8 Avg 78.8 74.9 81.2 77.8 Table 5: Detailed ablation results corresponding to Ta- ble 3 in the main paper across four audio chapterization subcategories (sc): Dynamic Media (DM), Gaming (G), Music (M), and Structured Speech (S). AF3-Think-8B is the unaligned base LALM, which lacks a sufficiently strong reasoning prior and does not reliably follow the target structured output format. AudioChaps-R1-Zero- 8B applies GRPO directly to AF3-Think-8B without an SFT cold start. AudioChaps-SFT-8B uses supervised fine-tuning to establish the target reasoning behavior and structured output format, while AudioChaps-R1-8B subsequently applies GRPO to calibrate the model’s chapter-boundary decisions. This ablation isolates the contribution of the SFT cold start and the subsequent GRPO alignment stage. The macro average (Avg) is reported over the four subcategories. est F1 on Dynamic Media, Gaming, Music, and Structured Speech. These results suggest that task-specific alignment can enable a compact open model to outperform substantially larger general- purpose LALMs on audio chapterization. This comparison evaluates the benefit of specialized alignment rather than the ultimate capability of stronger or larger backbones, which may improve further when adapted with AudioChaps. E Statistical Significance of the Main Improvement We assess the uncertainty of the performance dif- ference between AF3-Think-8B and AudioChaps- R1-8B using a paired non-parametric bootstrap. Because multiple evaluation clips may originate from the same source video and are therefore not ModelscAcc Pre RecF1 DM55.847.523.731.6 AF3-Think-8BG50.145.954.649.9 (Base-LALM)M62.364.03.26.0 S44.746.419.127.0 Avg53.251.025.228.6 DM75.970.077.273.4 AudioChaps-R1-8BG77.675.175.975.5 M88.183.685.684.6 S73.671.086.077.8 Avg78.874.981.277.8 DM 67.0 57.8 86.4 69.3 MOSS-Think-8BG70.9 63.3 85.9 72.9 (Base-LALM)M77.9 65.3 89.9 75.7 S70.4 66.7 89.6 76.5 Avg 71.5 63.3 88.0 73.6 DM 82.4 80.6 77.9 79.2 MOSS-Think-G80.8 80.6 75.9 78.2 AudioChaps-R1-8BM89.0 85.6 85.6 85.6 S81.5 82.7 82.9 82.8 Avg 83.4 82.4 80.6 81.4 Table 6: Backbone-generalization results across au- dio chapterization subcategories (sc): Dynamic Media (DM), Gaming (G), Music (M), and Structured Speech (S). Grey rows report the AF3-based models for refer- ence. Across both AF3 and MOSS, GRPO alignment improves precision (Pre), accuracy (Acc), and F1 while reducing the high-recall behaviour of the preceding model. Avg denotes the macro-average over subcate- gories. independent, we resample source videos rather than individual clips. In each of 10,000 bootstrap itera- tions, the same sampled videos are used to evalu- ate both models, preserving the paired comparison. For each resample, we compute positive-class F1 independently for each of the four acoustic subcat- egories and report their macro-average, matching the aggregation used in the main results. AudioChaps-R1-8B obtains a macro-F1 of 77.8, compared with 28.6 for AF3-Think-8B. The result- ing improvement is 49.2 F1 points, with a 95% bootstrap confidence interval of [46.2, 52.2] and a two-sided bootstrapp-value below10 −4 . The confidence interval excludes zero by a wide mar- gin. Improvements are also significant within every acoustic subtype: Dynamic Media,+41.8[37.5, 46.3]; Gaming,+25.6[19.9, 31.7]; Music,+78.6 [73.9, 83.0]; and Structured Speech,+50.8[42.4, 58.8]. The analysis includes 15,963 paired clips from 749 source videos. 13 F Full-Length Audio Chapterization Evaluation The clip-level AudioChaps-Eval benchmark pro- vides a controlled setting for isolating chapter- boundary judgment under balanced positive and negative sampling.We additionally evaluate whether these local decisions can be converted into recording-level chapter boundaries on uncropped, full-length audio. This setting is more challeng- ing because chapter boundaries are sparse, each recording may contain multiple boundaries, and overlapping window-level predictions must be con- solidated into a sequence of recording-level esti- mates. F.1 Evaluation Set and Sliding-Window Inference We evaluate all models on the common intersection of 40 full-length recordings, with 10 recordings from each of the four acoustic subcategories. The recordings have a mean duration of 49 minutes 29 seconds, yielding approximately 33 hours of long-form audio in total, and together contain 387 reference chapter boundaries. Each recording is divided into overlapping 60- second windows with a 20-second hop. A final tail window is added when necessary to ensure complete coverage of the recording. This produces 5,875 evaluation windows across the 40 recordings. F.2 Recording-Level Decoding For models that output only a binary boundary ver- dict, we use a window-run decoder. Consecutive windows predicted as positive are grouped into a contiguous run, and one boundary is placed at the midpoint of the run’s temporal coverage. Estimated boundaries separated by fewer than 20 seconds are subsequently merged by taking their median. This procedure prevents a single transition observed in several overlapping windows from being counted multiple times. F.3 Matching and Metrics Predicted and reference boundaries are matched one-to-one under a±10-second tolerance. This tol- erance follows from the training-data construction: positive 60-second crops place the reference bound- ary within the central 20 seconds, so a window- centre estimate localizes the boundary to within at most 10 seconds. Candidate matches within this tolerance are ordered by absolute temporal distance and greedily assigned without reusing either a pre- diction or a reference boundary. We compute precision, recall, and F1 indepen- dently for each recording, macro-average them within each acoustic subcategory, and then average equally across the four subcategories. We addition- ally reportdev-R2E, the pooled median distance from each reference boundary to its nearest pre- dicted boundary, and the ratio between the num- bers of predicted and reference boundaries. As a content-independent baseline, we place one bound- ary every 180 seconds. F.4 Results The fixed-interval baseline reaches only 9.5 aver- age F1, while AF3-Think-8B obtains 6.5 F1 and predicts no boundaries on 15 of the 40 record- ings. AudioChaps-R1-8B improves average F1 to 37.6 and reduces the pooled median reference-to- estimate deviation from 38 to 10 seconds, show- ing that clip-level training transfers to full-length recordings. With the window-run decoder, MOSS-Think- AudioChaps-R1-8B improves over the base LALM, MOSS-Think-8B, from 26.4 to 46.2 F1, primarily through a precision gain from 22.4 to 48.2. Overall, AudioChaps improves full-length chapterization across both backbones, demonstrating that locally trained boundary judgments can be consolidated into effective recording-level chapter predictions. G Training and Evaluation Dataset Statistics As discussed in Section 4.1, our data is curated from VidChapters-7M (Yang et al., 2023) and strat- ified into four acoustic subtypes. Here we provide a finer breakdown of the dataset by subcategory and by positive/negative clip balance. Figure 4 re- ports the per-subtype clip counts for the training and held-out test splits, together with the positive- to-negative ratio within each subtype. H Sample Output Comparison: AF3-Think vs. AudioChaps-R1 with Human Evaluation To assess reasoning quality beyond automatic met- rics, we conducted a blind human evaluation with seven raters, all PhD students or post-doctoral re- searchers. For each sample, a rater was shown a 60-second audio clip and the chapter-boundary question alongside two anonymized responses, one 14 from the AF3-Think base model and one from AudioChaps-R1, presented in randomized order to remove position bias. Each response was scored from 1 (poor) to 5 (excellent) on whether it cor- rectly identified the presence or absence of a bound- ary and on the quality of its justification. Averaged over all samples and raters, AF3-Think scored 2.77 while AudioChaps-R1 scored 4.46, a clear prefer- ence for the latter’s reasoning. The rater-facing instructions are reproduced in Figure 5, and repre- sentative samples with their ratings are shown in Figures 6, 7, 8, and 9. I Additional Temporal-Localization Analysis AudioChaps is designed to align the existing rea- soning capabilities of different LALM backbones with creator-authored chapter-boundary decisions. Most current LALMs do not provide sufficiently reliable timestamp predictions, so our primary task is to determine whether a chapter boundary oc- curs within an approximately±10-second interval, which supports practical chapter-level navigation in long-form media. More precise temporal local- ization may provide additional value, but it is not the central focus of this work. We therefore report timestamp-based experiments separately as a sec- ondary analysis for models that natively support temporal prediction. I.1 MOSS Timestamp Supervision MOSS-Think-8B has a comparatively strong native temporal-localization capability. AudioChaps can therefore exploit this capability in addition to align- ing its chapter-boundary decisions. Specifically, MOSS-Think-8B can emit an estimated bound- ary location within the 60-second input window. For the MOSS instantiation of AudioChaps, we include a small auxiliary timestamp reward, com- puted using the creator-annotated boundary times- tamp, alongside the boundary-classification objec- tive. This component leverages the backbone’s existing temporal capability and is not required by the primary AudioChaps framework. The AudioChaps-adapted MOSS model substan- tially improves temporal localization across all four subcategories. The mean absolute error (MAE) de- creases from 7.39 to 2.88 seconds on Dynamic Media, 9.42 to 4.49 seconds on Gaming, 5.47 to 2.46 seconds on Music, and 6.75 to 2.98 seconds on Structured Speech. These results show that the GRPO-based AudioChaps framework can lever- age the native temporal-localization capability of MOSS-Think-8B to achieve more accurate chapter- boundary localization. I.2 Timestamp-Aware Recording-Level Decoding In the full-length audio evaluation, we addition- ally examine whether MOSS’s in-window times- tamp predictions can improve the conversion of overlapping window-level outputs into recording- level chapter boundaries. For each positive predic- tion from a window beginning at absolute timet 0 , MOSS emits an in-window timestampt s . We con- vert this prediction into an absolute boundary vote att 0 + t s . Votes are grouped using single-linkage clustering with a 10-second linkage threshold, and each cluster produces one boundary estimate at the median of its member predictions. Predicted timestamps outside the 60-second input interval are discarded. This timestamp-based decoder is evaluated sep- arately from the window-run decoder used in the primary full-length comparison. For MOSS-Think- AudioChaps-R1-8B, it raises precision from 48.2 to 54.3, recall from 49.3 to 64.9, and F1 from 46.2 to 55.2. It also reduces the pooled median reference- to-estimate deviation from 8 to 2 seconds. Among timestamp clusters supported by multiple overlap- ping windows, the median within-cluster spread is 1 second and the 90th-percentile spread is 8 sec- onds. Overall, timestamp-aware decoding provides ad- ditional gains in the full-length setting when the un- derlying LALM supports reliable temporal outputs. These results complement the primary window-run evaluation but do not alter the central focus of Au- dioChaps on aligning chapter-boundary judgments with creator-authored editorial annotations. 15 ModelscAccPreRecF1 DM 72.5 66.7 78.8 72.2 Gemini-2.5-FlashG73.0 67.9 82.5 74.5 (Zero-Shot)M81.0 74.9 81.4 78.0 S78.3 77.3 88.0 82.3 Avg 76.2 71.7 82.7 76.8 DM 73.1 68.6 69.2 68.9 Qwen3-Omni-30B-G75.6 74.9 69.7 72.2 A3B-ThinkingM85.3 80.6 81.0 80.8 (Zero-Shot)S77.6 78.7 79.8 79.2 Avg 77.9 75.7 74.9 75.3 DM 67.0 57.8 86.4 69.3 MOSS-Think-8BG70.9 63.3 85.9 72.9 (Zero-Shot)M77.9 65.3 89.9 75.7 S70.4 66.7 89.6 76.5 Avg 71.5 63.3 88.0 73.6 DM 60.2 55.3 62.7 58.8 Step-Audio-R1-32BG60.1 59.1 53.4 56.1 (Zero-Shot)M62.4 53.1 51.9 52.5 S65.6 70.1 69.7 69.9 Avg 62.1 59.4 59.4 59.3 DM 55.8 47.5 23.7 31.6 AF3-Think-8BG50.1 45.9 54.6 49.9 (Zero-Shot)M62.3 64.03.26.0 S44.7 46.4 19.1 27.0 Avg 53.2 51.0 25.2 28.6 DM 75.9 70.0 77.2 73.4 AudioChaps-R1-8BG77.6 75.1 75.9 75.5 M88.1 83.6 85.6 84.6 S73.6 71.0 86.0 77.8 Avg 78.8 74.9 81.2 77.8 DM 82.4 80.6 77.9 79.2 MOSS-Think-G80.8 80.6 75.9 78.2 AudioChaps-R1-8BM89.0 85.6 85.6 85.6 S81.5 82.7 82.9 82.8 Avg 83.4 82.4 80.6 81.4 Table 7: Comparison with additional multimodal and omni-modal LLMs on AudioChaps-Eval across four audio chapterization subcategories: Dynamic Media (DM), Gaming (G), Music (M), and Structured Speech (S). Accuracy (Acc), precision (Pre), recall (Rec), and F1 are reported as percentages, together with the macro- average (Avg) over the four subcategories. Gemini- 2.5-Flash, Qwen3-Omni-30B-A3B-Thinking, MOSS- Think-8B, Step-Audio-R1-32B, and AF3-Think-8B are evaluated in the zero-shot setting. The two AudioChaps- R1 variants apply our alignment framework to the AF3- Think and MOSS-Think backbones, respectively. ModelscPreRecF1dev-R2E↓ DM18.9 13.5 10.7– AF3-Think-8BG3.410.54.6– (Window)M0.41.70.7– S11.8 12.7 10.0– Overall 8.6 9.6 6.5 38.0 s DM32.6 38.6 34.5– AudioChaps-R1-8BG20.0 38.3 24.6– (Window)M58.1 73.7 63.7– S25.8 32.3 27.5– Overall 34.1 45.7 37.6 10.0 s DM17.4 26.7 20.4– MOSS-Think-8BG6.519.98.1– (Window)M35.8 63.4 44.2– S30.1 38.5 32.8– Overall 22.4 37.1 26.4 15.0 s DM53.3 41.1 42.4– MOSS-Think-G26.2 27.4 23.2– AudioChaps-R1-8BM60.4 73.0 65.5– (Window)S52.6 55.5 53.7– Overall 48.2 49.3 46.2 8.0 s DM12.0 13.8 12.5– Fixed IntervalG8.88.68.2– (180 s)M9.415.1 11.0– S5.98.76.2– Overall 9.0 11.6 9.5 49.0 s Table 8: Full-length audio chapter-boundary detection across four acoustic subcategories: Dynamic Media (DM), Gaming (G), Music (M), and Structured Speech (S). Predictions are matched one-to-one with reference boundaries under a±10-second tolerance. Precision (Pre), recall (Rec), and F1 are computed per recording, macro-averaged within each subcategory, and then av- eraged equally across the four subcategories.dev-R2E denotes the pooled median distance from each reference boundary to its nearest prediction and is reported only in the Overall row; lower values indicate better local- ization. The window-run decoder merges contiguous positive windows into a single recording-level boundary estimate. 16 Figure 4: Per-subtype clip counts for the training (AudioChaps-Alignment) and held-out test (AudioChaps-Eval) splits, split by positive (boundary-spanning) and negative (within-chapter) clips. Both splits follow the same four-way stratification, with no source-video overlap between the training set and the held-out test set. Instructions shown to raters Study description. This study evaluates an LALM trained to detect structural chapter boundaries in YouTube video clips (we consider audio only). For each sample, the models are provided with a 1-minute audio clip and a targeted prompt. You will see two randomized model responses (Output A and Output B). Please listen to the clip, review the scripts, and rate each model’s ability to accurately identify whether a boundary exists, as well as the quality of its justification. Important note. Model outputs (A, B) are randomized per sample to prevent evaluation bias. For example, Model 1’s output might appear as Output A in the first sample and Output B in the next. Figure 5: Instructions shown to human evaluators. Each sample pairs a 1-minute clip and the boundary question with two anonymized, randomly ordered model outputs (Output A and Output B), each rated on a 1–5 scale. 0:00/3:530:00/3:53 AF3 Based on the audio, the track maintains a consistent musical style and structure throughout its duration, without any significant changes or distinct sections that would indicate a 'chapter' boundary. Therefore, the answer is: **(B) No** HE : 1.86 AudioChaps <think>Let's break down the audio flow. First, the audio is a high-energy electronic track. It's characterized by a fast tempo, a driving beat, and a repetitive, robotic vocal sample saying \"Found\". The overall feeling is chaotic and upbeat. , this music cuts off abruptly. This is followed by a period of complete silence. that, a completely different style of music begins. This new track is much slower, with a steady, pulsing beat and a much more melodic and relaxed feel. The vocal sample changes from a robotic \"Found\" to a clear, human-sounding \"F-O-U-R\" spelled out letter by letter. The entire atmosphere shifts from chaotic and energetic to calm and deliberate. transition from one distinct musical piece to a completely different one, separated by a period of silence, is a very strong indicator of a chapter boundary. The first activity (the fast song) ends, and a new one (the slow song) begins.</think> <answer>A</answer> HE : 4.57 Figure 6: Qualitative comparison of AF3-Think-8B and AudioChaps-R1-8B on human-evaluation (HE). 17 0:00/3:530:00/3:53 AF3 Based on the audio, the speaker is providing a live commentary on a race, mentioning competitors like Minty Drizzle, Momo, Yellow Eye, and Razi, and describing their positions and actions. This type of continuous, real-time reporting is characteristic of a race broadcast, which typically includes a chapter boundary to segment the content. Therefore, the presence of a chapter boundary is implied by the nature of the audio content. The correct answer is: **(A) Yes** HE : 2.86 AudioChaps <think>Let's break down the audio flow. The clip starts with a commentator announcing the elimination of a participant. This is followed by a shift to a race update, where the commentator describes the positions of the remaining competitors. The narration then becomes more detailed, focusing on a specific driver's performance and the race's progress. The commentator continues to provide updates on the race, mentioning changes in lead and positions. The entire clip seems to be a single, continuous commentary on a race. The topic remains consistent throughout, and the speaker's tone and pace are also consistent. The clip ends with the commentator still providing race updates. There is no shift to a new topic, no change in the overall activity, and no transition to a different segment. The clip remains within the same activity from start to finish.</think> <answer>B</answer> HE : 4.57 Figure 7: Qualitative comparison of AF3-Think-8B and AudioChaps-R1-8B on human-evaluation. 0:00/3:530:00/3:53 AF3 Based on the audio, the speaker explicitly states \"This is the end of the first chapter\" around the 00:00 mark. This clearly indicates a chapter boundary is present in the audio. Therefore, the correct answer is: **(A) Yes** HE : 1.29 AudioChaps <think>Let's break down the audio flow. The clip starts with a speaker commentating on a video game, describing the difficulty of using a weapon and missing shots. This is followed by the speaker reacting to the game's events, like seeing a pile of bodies and a character who is alive. The speaker's tone shifts from calm to more urgent as the action intensifies, with sounds of gunfire and explosions. The speaker then seems to strategize, saying they know what to do. This is followed by a brief dialogue with a female character, who is determined to find someone named Daisy. The speaker then expresses a conflict, stating he doesn't want to hurt Daisy but wants to find her. The clip ends with the speaker's voice rising in urgency and the sound of a weapon being readied. The entire sequence is a continuous stream of commentary and in-game action. The speaker is engaged in a single, uninterrupted gameplay session. The changes in tone and the dialogue with the character are all part of this one activity. The clip remains within the same activity from start to finish.</think> <answer>B</answer> HE : 4.43 Figure 8: Qualitative comparison of AF3-Think-8B and AudioChaps-R1-8B on human-evaluation. 18 0:00/3:530:00/3:53 AF3 Based on the audio, the music maintains a consistent high-energy, aggressive sound throughout the clip, without any noticeable shift in structure or intensity that would indicate a distinct \"chapter\" or section. Therefore, the answer is: **(B) No** HE : 1.86 AudioChaps <think>Let's break down the audio flow. First, the audio is filled with the intense sound of a heavy metal song, complete with distorted guitars, fast drumming, and a loud, cheering crowd. This creates a very specific, high-energy atmosphere. Then, this music and crowd noise come to a sudden and complete stop. This is followed by a period of silence, which is then broken by a speaker's voice. The speaker's tone is completely different from the music; it's calm, conversational, and announces the next song. After that, a completely new musical piece begins. This new song has a different style, with cleaner guitars and a slower tempo. The crowd's reaction also changes, becoming more subdued. The shift from one distinct song and its conclusion, through a spoken interlude, to a completely new song is a very clear chapter boundary. The first activity (the first song) ends, and a new one (the second song) begins.</think> <answer>A</answer> HE : 4.71 Figure 9: Qualitative comparison of AF3-Think-8B and AudioChaps-R1-8B on human-evaluation. 19