Discovering and Causally Validating Emotion-Sensitive Neurons in Large Audio-Language Models

AI summary

Overview

Research area: Mechanistic interpretability of large audio-language models (LALMs), specifically neuron-level analysis of how these models encode emotion in speech.

Technical level: Advanced. The paper assumes familiarity with transformer internals (decoder MLP blocks, SwiGLU gating), inference-time activation interventions, and speech emotion recognition (SER) benchmarks.

Scope: The paper identifies and causally validates emotion-sensitive neurons (ESNs) in three open-source LALMs — Qwen2.5-Omni-7B, Kimi-Audio, and Audio Flamingo 3 — across three SER datasets, comparing five neuron-selection methods and testing deactivation, targeted steering, and label-free injection.

What This Paper Is About

Large audio-language models handle speech input and often appear to "understand" a speaker's emotional state, but no one has shown how that emotion information is stored inside them. This paper asks whether these models contain a compact set of individual neurons that preferentially fire for particular emotions, and whether turning those neurons off or amplifying them selectively changes the model's emotion judgments. The authors call these units emotion-sensitive neurons (ESNs) and set out to test them causally rather than just correlationally.

Key Contributions

1. First neuron-level causal analysis of emotion representations in LALMs. The authors state this is, to their knowledge, the first study to identify emotion-related neurons in large audio-language models and validate them with self/cross deactivation and steering interventions across multiple models and datasets.

2. A systematic comparison of neuron-identification criteria. Five selectors are compared — Random (RND), Activation Probability (LAP), Activation Probability Entropy (LAPE), Mean Activation Difference (MAD), and Contrastive Activation Margin (ConAct) — showing which criteria isolate causally emotion-sensitive units and how selection parameters matter.

3. Evidence of structured locality and non-trivial transfer. ESNs are shown to cluster non-uniformly across decoder layers and to transfer partially, but unevenly, across datasets.

4. Demonstration of targeted and label-free control. Amplifying ESNs provides a controllable handle for biasing predictions toward a target emotion, while three label-free (agnostic) injection strategies — 2-Pass, Mix, and Union — are evaluated as weaker, less consistent alternatives.

Main Findings

• Deactivation produces an emotion-specific signature. Masks built from MAD- and ConAct-selected neurons cause large negative self-effects (the paper reports 11–15 accuracy points across the three models) while leaving average cross-emotion effects near zero, yielding substantial self–cross gaps. Qwen2.5-Omni-7B: MAD self-effect −13.09, ConAct −13.50, with self–cross gaps of −13.28 and −15.25. Kimi-Audio: MAD −13.63 and ConAct −11.65, gaps of −12.36 and −12.09. Audio Flamingo 3: MAD −15.17 and ConAct −14.63, gaps of −13.21 and −15.33.

• Not all selectors work equally well. LAP and LAPE often fail to produce a clean diagonal signature, yielding weak, noisy, or broader degradation instead. Random masks of the same size (averaged over 5 independent random masks of different seeds) produce smaller and less structured changes, supporting the claim that the effects are not merely generic capacity loss.

• Targeted steering is constructive. Amplifying the same ESN sets produces consistent self-steering gains of roughly +2 to +3 accuracy points while leaving cross-steering effects largely unchanged on average. Table 1 values include Qwen2.5-Omni-7B MAD 2.48 and ConAct 2.73; Kimi-Audio MAD 2.25 and ConAct 1.94; Audio Flamingo 3 MAD 2.97 and ConAct 3.35 (self–cross gaps up to 4.07).

• Intervention size trades selectivity for strength. At small fractions (r = 0.1%), deactivation already yields clear diagonal patterns; as r grows to 0.5% and 1.0%, self-deactivation strengthens but off-diagonal spillover grows. The authors settle on r = 0.5% to balance causal potency against clean self–cross dissociation. Steering shows a parallel trade-off as the gain α increases (tested at 0.10, 0.30, 0.50, and 1.00).

• Modest identification data suffices. The accuracy curves plateau rapidly as the number of correctly answered identification examples per emotion increases, with larger pools yielding diminishing returns — so stable neuron identification does not require extremely large pools.

• Agnostic injection is weaker and model-dependent. Mix and Union improve Qwen2.5-Omni-7B (up to +0.9 for Mix) and Audio Flamingo 3 (up to +1.0 for Union), but all strategies slightly underperform the unmasked baseline on Kimi-Audio. The authors interpret the gap versus targeted steering as evidence that ESNs may interact non-additively under joint amplification.

• ESNs are non-uniformly distributed by layer. All three models have 28-layer decoders. ESNs cluster in the earliest layer (layer 0), early–mid layers (6–8), and later layers (19–22), with relatively sparse presence in central blocks (15–18). The Neutral category shows the strongest emotion-specific deviations.

• Cross-dataset transfer is partial and asymmetric. Across all six transfer directions, recurring diagonal structure appears for shared emotions, indicating that many ESNs encode dataset-robust affective computations rather than corpus-specific artifacts. Transfer strength is uneven and sometimes asymmetric. Neutral shows the least stable transfer, often producing smaller or non-diagonal effects.

Methodology in Plain English

The authors follow an activation-based log–identify–intervene workflow, using speech emotion recognition as a diagnostic task.

First, they attach forward hooks to the decoder MLP feed-forward blocks of each model and record the gating signal from the SwiGLU nonlinearity. They log activations only on items the unintervened model answered correctly, to reduce contamination from failure-mode generations and get cleaner emotion-conditioned statistics.

Second, they compute, for each neuron and emotion, how often it fires (the fraction of valid token positions with a positive activation). From this firing-probability profile they score neurons with five different criteria: random selection as an emotion-agnostic control; LAP, which favors frequently firing neurons; LAPE, which favors neurons whose firing distribution over emotions has low Shannon entropy; MAD, which measures how far a neuron's firing probability for one emotion deviates from its average across all emotions; and ConAct, a margin-style criterion comparing the top firing probability across emotions with the runner-up. For each method and emotion they take a fixed fraction r of the highest-ranked neurons as the ESN set.

Third, they intervene at inference time. Deactivation zeroes out the selected gate activations; targeted steering multiplies them by a gain of 1 + α. Neither changes any model weights. They then compare a self-effect (neurons selected for the same emotion being evaluated) against a cross-effect (an emotion-e mask evaluated on a different emotion subset), and quantify specificity as the self–cross gap.

Three agnostic injection strategies are also tested: 2-Pass, which reinforces the model's own first-pass prediction; Mix, which softly weights emotion masks using internal evidence; and Union, which boosts the union of all ESNs.

Evaluation uses a controlled multiple-choice format with a single instruction template, randomized option-number-to-emotion mapping on every item to prevent position preferences from inflating any one emotion, greedy decoding at temperature 0, and a 20-token generation cap. Outputs are normalized to a single option index through a cascade of numeric parsing, spelled-out number mapping, and fallback emotion-string matching.

Why This Matters

Impact on research. The paper moves LALM interpretability from correlation to causation in the affective domain. Prior multimodal neuron-level work focused largely on modality- or task-specific patterns rather than emotion, and on LLMs rather than speech-enabled models. By showing that manipulating a tiny fraction of neurons (0.5%) produces an emotion-specific causal signature in three independent architectures, the work offers a template for how to validate interpretability claims and provides a mechanistic counterpart to existing behavioral SER evaluations. It also connects to prior work on clustered emotion neurons and emotion circuits in text-only LLMs, extending that line into audio-language models.

Real-world applications:

• Conversational assistants and voice agents. Affective competence is described in the paper as crucial for user trust and safety; a neuron-level handle could allow a system to be tuned toward or away from particular affective responses.
• Speech emotion recognition systems. Understanding which internal units drive emotion decisions could inform more robust or more auditable SER deployments.
• Expressive speech synthesis. The authors note expressive speech synthesis as an adjacent task and flag cross-task generalization as future work, so control over affective representations is a plausible downstream application.
• Model auditing and debugging. Locating emotion decisions in specific layers (0, 6–8, 19–22) gives engineers concrete intervention points rather than a black box.

Industry relevance. The results are directly usable by teams building or deploying speech-enabled foundation models, since the interventions require no retraining, no weight modification, and only a modest identification pool of correctly answered examples. The method-dependence finding is equally practical: practitioners who reach for a simple activation-frequency or entropy criterion (LAP/LAPE) will get weak, noisy, and often non-selective neuron sets, whereas MAD- or ConAct-style deviation and margin criteria yield compact sets with strong causal effects.

Future Directions

1. Characterizing interactions between ESNs. The weaker, model-dependent results of agnostic injection hint at non-additive or competitive interaction among ESNs. The authors explicitly state they do not yet provide a dedicated causal decomposition of these interactions, and call for pairwise co-steering or controlled multi-emotion activation studies to map competitive versus cooperative affective circuitry.

2. Extending beyond the decoder MLP gates. The interventions cover only decoder SwiGLU gate activations. The limitations section notes this does not characterize how emotion cues are distributed across attention and audio–text fusion components, nor how multiple units compose into higher-level circuits.

3. Generalization beyond SER. Transfer in this study is measured across datasets within the emotion recognition setting. Whether ESNs generalize across tasks such as expressive speech synthesis, and how to make steering uniformly reliable, remain open questions.

4. Improving label-free control. Since targeted steering requires specifying a source emotion, making agnostic injection reliable without committing to a target emotion is an unresolved practical problem, and one that would make neuron-level control usable in settings where the desired emotion label is unknown.

Target Audience

This paper is aimed at interpretability researchers working on multimodal and speech-enabled foundation models, and at speech and affective computing researchers interested in the mechanisms behind SER. It will also be useful to machine learning engineers who want a concrete, cheap intervention technique for controlling affective behavior in audio-language models, and to graduate students already comfortable with transformer internals who want a methodological template for moving from neuron identification to causal validation. Readers without background in activation interventions or SwiGLU gating will find the paper heavy going.

Authors’ abstract

Emotion is a central dimension of spoken communication, yet, we still lack a mechanistic account of how modern large audio-language models (LALMs) encode it internally. We present the first neuron-level interpretability study of emotion-sensitive neurons (ESNs) in LALMs and provide causal evidence supporting the existence of such units in Qwen2.5-Omni, Kimi-Audio, and Audio Flamingo 3. Across these three widely used open-source models, we compare frequency-, entropy-, mean-deviation-, and contrast-based neuron selectors on multiple emotion recognition benchmarks. Using inference-time interventions, we reveal a consistent emotion-specific signature: deactivating neurons selected for a given emotion disproportionately degrades recognition of that emotion while largely preserving other classes, whereas targeted steering amplifies these units to bias predictions toward the target emotion. These effects arise with modest amounts of identification data and scale systematically with intervention strength. We further observe that ESNs exhibit non-uniform layer-wise clustering with partial cross-dataset transfer. Taken together, our results offer a causal, neuron-level account of emotion decisions in LALMs and highlight targeted neuron interventions as an actionable handle for controllable affective behaviors.