When Does the Concept of "Dog" Emerge in an Audio LLM?

1. INTRODUCTION

• The study addresses how nonlinguistic sounds enter verbalizable representations and contribute to responses across tasks, which final answers alone cannot reveal.Readout tests decodability, while interventions test influence on outputs; score changes may not alter final answers, motivating full-response generation and controls.Source: §1 · p. 1
• The authors study dog barking in Qwen2.5-Omni-7B using a J-lens-derived dog direction and compare dog, bark, and animal concepts, with cat vocalizations as a related comparison.Source: §1 · p. 1
• Dog-related information is decodable without dog/bark prompt cues or animal-identification requirements.Source: §1 · p. 2
• For non-dog inputs, dog-direction injection produces higher answer-transition rates than norm-matched controls at the compared dose.The authors present this as causal evidence that the direction influences answer generation.Source: §1 · p. 2
• The dog direction most consistently affects responses at L22–L24 before generation across species, vocalization, and sound-description tasks, but not animal/other judgments.Removing or adding the dog component respectively lowers or raises dog-related answer scores.Source: §1 · p. 2

2.1. Model and Data

• The Thinker language decoder has 28 Transformer blocks, and recordings come from DogSpeak, an assembled cat-vocalization dataset, ESC-50, and UrbanSound8K.Concept ranking uses 50 dog and 50 non-dog recordings; readout and intervention use 600 recordings split evenly among dog, cat, and other sounds.Source: §2.1 · p. 2

2.2. J-lens Mapping and Concept Readout

• J-lens maps selected layer states to L26, and concept readout uses output-head token vectors to assess dog/non-dog decodability.The mappings are text-fitted and validated on held-out prompts and audio without audio-domain refitting; full-vocabulary ranks and AUCs are also used.Source: §2.2 · p. 2

2.3. Intervention Methods and Controls

• Interventions test dog ablation, injection, and dog–cat swapping against controls, using normalized directions for ablation and injection and unnormalized directions for swapping.Vehicle and random orthogonal directions serve as controls; generation and localization strictly match perturbation norms.Source: §2.3 · pp. 2–3

2.4. Tasks and Experimental Procedure

• The study compares dog-related readout and interventions across seven tasks, including species, vocalization type, sound description, and animal/other classification.Fixed-response, loudness, and event-count tasks do not require animal identification; the first three tasks omit dog/bark from prompts.Source: §2.4 · p. 3

2.5. Evaluation Metrics and Statistical Tests

• Evaluation measures readout with AUC and score differences, and measures intervention effects with first-step logits or task-specific log-odds.Localization compares dog-direction changes with controls; support requires the expected score direction, a confidence interval above zero, and q < 0.05 after correction.Source: §2.5 · p. 3, Table 1 · p. 3

3.1. Mapping Validation and Concept Readout

• Audio mapping validation reaches cosine similarity 0.820 at L24 with 98.5% output-head top-1 agreement, while substantial mapping error limits early-layer interpretation.Source: §3.1 · p. 3
• At L24, dog distinguishes dog-barking from non-dog recordings with AUC 0.937, while bark remains informative with AUC 0.872 despite ranking far below top vocabulary items.The combined score (sdog + sbark) − (ssound + snoise) has higher AUC than sbark − snoise in this experiment.Source: §3.1 · p. 3

3.2. Concept Readout in Tasks Unrelated to Animal Identification

• Dog and bark remain discriminative under fixed-response and loudness prompts that contain neither dog nor bark and do not require animal identification.Table 2 reports separability for selected tasks; its caption defines ∆read as the mean readout-score difference.Source: Table 2 · p. 4, §3.2 · p. 4

3.3. Effects of Directional Interventions on Output Scores and Final Generated Answers

• At L24, ablation decreases the dog logit by 2.806 on average, injection increases it by 1.279, and swapping changes the dog–cat logit difference by −4.799 for dog samples and +4.396 for cat samples.Effects are close to zero at L0 and L8, increase from L20 onward, and are most pronounced at L24.Source: §3.3 · p. 4
• At L24, β=2 under label order A yields 53/191 dog→non-dog transitions versus 14/191 at β=1; α=0.06 injection yields 20/180 other→dog versus 0/180 vehicle and 1/180 random controls.Swapping at γ = 1.5 yields 93/191 dog→cat transitions and 1/190 cat→dog transitions.Source: §3.3 · p. 4

3.4. Intervention Localization and Task Differences

• At L22 and L24, six of eight task–operation comparisons meet support criteria, versus at most two at other tested positions; animal/other classification shows no corresponding advantage.Limited baseline performance in vocalization classification and sound description means cross-task conclusions rely mainly on continuous scores; final-answer evidence mainly comes from species classification.Source: Table 3 · p. 4, §3.4 · p. 4

• Dog-related information remains decodable without dog/bark prompt cues or animal-identification requirements.Source: §4 · p. 4
• Dog-direction interventions change output scores and some final species-classification answers, with effects most consistent at L22 and L24 immediately before answer generation.Source: §4 · p. 4
• The direction shows task-dependent effects: larger score changes than matched controls in species classification, vocalization classification, and sound description, but no corresponding advantage in animal/other judgments.Source: §4 · p. 4