Paper deep dive
Tracing a Sparse Emotion-Control Circuit in LLM-Based Text-to-Speech
Hongfei Du, Jiacheng Shi, Yanfu Zhang, Ye Gao
Intelligence
Status: not_run | Model: - | Prompt: - | Confidence: 0%
Entities (0)
Relation Signals (0)
No relation signals yet.
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:LLM-based text-to-speech (TTS) models can generate emotionally expressive speech, but how reference emotion is routed through the model and realized in decoded speech remains unclear. We introduce two emotion-sensitive metrics for matched neutral and emotional syntheses---a codec trajectory score and a late residual direction score---and use them to score activation-patching interventions. Under controlled matched-reference conditions, this analysis identifies a sparse source-to-readout component-level circuit: 23--27 attention heads and MLPs per emotion, roughly 5% of the components considered, recover or suppress 74--88% of the late emotion-readout shift on held-out cases. The circuit combines a shared component backbone with emotion-specific components; cross-emotion activation swaps reduce the target readout in 47 of 48 cases. In decoded speech, the same intervention produces consistent changes in pitch, energy, and spectral brightness over 24 matched pairs per emotion. A readout-matched residual-direction baseline produces only 17--27% of the intervention's pitch effect, showing that internal readout movement alone does not explain the decoded acoustic changes. These results trace a compact causal route from reference-derived prefix information to emotion-relevant properties of generated speech.
Tags
Links
- Source: https://arxiv.org/abs/2610.05080v1
- Canonical: https://arxiv.org/abs/2610.05080v1
Trouble viewing inline? Open PDF directly →
Full Text
60,528 characters extracted from source content.
Expand or collapse full text
Tracing a Sparse Emotion-Control Circuit in LLM-Based Text-to-Speech Hongfei Du Jiacheng Shi Yanfu Zhang Ye Gao Affiliation: Department of Computer Science, William & Mary, Williamsburg, USA Affiliation: hdu02, jshi12, yzhang105, ygao18@wm.edu Abstract LLM-based text-to-speech (TTS) models can generate emotionally expressive speech, but how reference emotion is routed through the model and realized in decoded speech remains unclear. We introduce two emotion-sensitive metrics for matched neutral and emotional syntheses—a codec trajectory score and a late residual direction score—and use them to score activation-patching interventions. Under controlled matched-reference conditions, this analysis identifies a sparse source-to-readout component-level circuit: 23–27 attention heads and MLPs per emotion, roughly 5% of the components considered, recover or suppress 74–88% of the late emotion-readout shift on held-out cases. The circuit combines a shared component backbone with emotion-specific components; cross-emotion activation swaps reduce the target readout in 47 of 48 cases. In decoded speech, the same intervention produces consistent changes in pitch, energy, and spectral brightness over 24 matched pairs per emotion. A readout-matched residual-direction baseline produces only 17–27% of the intervention’s pitch effect, showing that internal readout movement alone does not explain the decoded acoustic changes. These results trace a compact causal route from reference-derived prefix information to emotion-relevant properties of generated speech. 1 Introduction LLM-based text-to-speech (TTS) models have recently shown strong capability for generating natural and emotionally expressive speech from text, speaker information, and reference prompts (Wang et al., 2023a; Du et al., 2024; Wang et al., 2025; Deng et al., 2025; Zhou et al., 2026). Controllable style and emotion generation is especially important for spoken dialogue agents, virtual humans, and health-support interfaces (Picard, 1997; Bickmore and Picard, 2005), because vocal emotion and prosody shape how listeners perceive intent, engagement, and empathy (Scherer, 2003; Juslin and Laukka, 2003). Recent TTS systems support such control through reference prompts, style descriptions, or explicit emotion conditioning (Guo et al., 2023; Yang et al., 2024; Zhou et al., 2026; Yang et al., 2025). This capability raises a mechanistic question: how does reference emotion enter an LLM-based TTS model, and which internal components carry its effect toward generated speech? Prior mechanistic studies of language models often study behaviors with localized output positions (Olah et al., 2020; Elhage et al., 2021). In factual recall, causal tracing, and automated circuit discovery, interventions can often be scored at answer tokens or other local prediction sites (Vig et al., 2020; Meng et al., 2022; Heimersheim and Nanda, 2024; Conmy et al., 2023). This scoring strategy does not transfer directly to LLM-based TTS. Reference emotion affects a long sequence of generated speech tokens, which are then transformed by acoustic decoding before they become audible. A decoded-speech emotion score can verify that decoded speech moves toward a target emotion, but it does not reveal which hidden states, components, or generation steps caused the change. Mechanistic analysis of TTS therefore needs measurements that link internal activations to codec-level changes and decoded-speech outcomes. Our causal tracing workflow has four steps. We first build matched neutral and emotional synthesis pairs, then localize where reference emotion enters the conditioning prefix, trace its effect through attention heads and MLPs to late emotion readouts, and finally test whether the selected components causally change decoded speech. Together, these steps turn a decoded-speech emotion change into a source-to-readout causal localization problem. Figure 1: Causal tracing localizes a sparse source-to-readout component-level emotion circuit within a reference-conditioned LLM-based TTS model. Matched neutral/emotional runs localize a conditioning-prefix source, trace its effect through sparse attention-head and MLP sets to late emotion readouts, and validate the pathway with causal and decoded-speech tests. To causally localize the reference emotion effect, we construct matched neutral and emotional syntheses that keep text and speaker setting fixed while changing the emotion reference. For these pairs, we design and validate two emotion-sensitive metrics: a codec trajectory score over generated token trajectories and a late residual direction score over generation states. These metrics let us score activation-patching interventions that replace selected internal activations from one synthesis trajectory with matched activations from another. These scores support prefix localization, component-level tracing, matched controls, component specificity checks, and decoded-audio validation. Figure 1 summarizes this complete workflow. Our activation-patching analysis identifies a sparse source-to-readout component-level circuit under controlled matched-reference conditions. The circuit starts at a shared conditioning-prefix source and passes through 23–27 selected attention heads and MLPs per emotion, roughly 5% of the components considered, to late emotion readouts. On held-out cases, these components recover or suppress 74–88% of the readout shift and exceed matched random sets in every emotion and intervention direction. The circuit has a shared computational backbone together with emotion-specific components: shared-only sets recover 79–86% of the full-set effect, while cross-emotion donor activations reduce the target readout in 47 of 48 cases. Finally, paired analyses over 24 decoded waveforms per emotion show consistent changes in pitch, energy, and spectral brightness. A residual-direction baseline matched to the internal readout shift produces only 17–27% of the selected components’ pitch effect, connecting the localized components to acoustic changes beyond a global emotion direction. Our main contributions are: • We design and validate two emotion-sensitive metrics for causal localization in TTS: a codec trajectory score and a late residual direction score. Unlike prior circuit work that scores answer-token interventions, these metrics score patching across codec-generation trajectories and late residual states, enabling localization that can be tested against decoded waveform evidence. • We identify a sparse source-to-readout component-level circuit for reference emotion. It begins at a shared conditioning-prefix source and passes through 23–27 selected attention heads and MLPs per emotion to late readouts for anger, happiness, and sadness. • We show that the circuit is causal and structured. Sufficiency and necessity interventions recover or suppress 74–88% of the readout shift; shared/specific decomposition and cross-emotion swaps reveal a shared backbone carrying emotion-dependent activation content. • We connect the internal circuit to decoded speech. The selected components produce consistent emotion-relevant acoustic changes and substantially outperform readout-matched residual-direction steering at transferring the internal shift to waveform pitch. 2 Related Work Emotion control in TTS. Emotion-controllable TTS aims to synthesize speech whose affect can be specified or guided by the user. Existing systems expose this control through natural-language style descriptions, explicit emotion labels, intensity controls, reference prompts, or reinforcement-learning objectives (Guo et al., 2023; Yang et al., 2024; Zhang et al., 2025; Yang et al., 2025; Li et al., 2026). Modern TTS systems often synthesize speech by generating discrete semantic or acoustic tokens before waveform decoding (Borsos et al., 2023; Wang et al., 2023a). Recent LLM-based TTS systems, including CosyVoice, Spark-TTS, IndexTTS, and IndexTTS2, build on this token-generation design and support controllable generation through attribute prompts, style references, or emotion inputs (Du et al., 2024; Wang et al., 2025; Deng et al., 2025; Zhou et al., 2026). These systems demonstrate practical emotion control in generated speech, but they do not identify how reference-conditioned emotion is represented or routed inside the TTS model. We instead ask where reference-conditioned emotion enters LLM-based TTS generation and which internal components carry its effect to emotion-sensitive readouts. Causal circuit localization. Mechanistic interpretability studies often explain a model behavior by identifying the internal states or components that causally support a chosen metric (Olah et al., 2020; Elhage et al., 2021; Olsson et al., 2022). Causal mediation, causal tracing, and activation patching test this relation by replacing internal states under a source-target contrast (Vig et al., 2020; Meng et al., 2022; Heimersheim and Nanda, 2024). Circuit studies then refine this localization from residual states to attention heads, MLPs, or paths (Wang et al., 2023b; Conmy et al., 2023; Goldowsky-Dill et al., 2023; Kramár et al., 2024). Most circuit-level evidence so far comes from localized prediction tasks, such as factual recall or indirect object identification. We study reference-conditioned emotion in LLM-based TTS under this causal localization perspective. Unlike localized prediction tasks, the emotion effect is distributed across a long sequence of generated speech tokens and only becomes audible after waveform decoding. We therefore need intervention metrics that score how changes in prefix and generation states affect the generated token trajectory and decoded speech. Internal analyses of speech models. Recent work has begun to analyze internal representations in speech models. Mechanistic studies of automatic speech recognition examine how speech representations support transcription-related behavior (Glazer et al., 2026). Paek et al. (2025) analyze audio latent spaces with sparse autoencoders and learn features aligned with acoustic properties such as pitch, amplitude, and timbre. In speech synthesis, hidden-state steering methods show that adding learned directions can change the emotion of synthesized speech without updating model weights (Xie et al., 2025). These studies show that speech model internals can be probed or manipulated, but they do not trace how reference-conditioned emotion flows through an LLM-based TTS model. We instead localize a conditioning-prefix source and identify attention heads and MLPs that carry its effect to late emotion readouts. 3 Preliminaries Generated speech token trajectories. We study the autoregressive stage that produces generated speech tokens before waveform decoding. In LLM-based TTS, an autoregressive language-model component predicts discrete speech tokens from text, speaker information, and an emotion reference, and a downstream decoder converts the predicted token sequence into a waveform (Wang et al., 2023a; Du et al., 2024; Wang et al., 2025; Deng et al., 2025). Suppressing speaker notation for clarity, we write this process as pθ(zt∣x,r,z<t),y=D(z1:T),p_θ(z_t x,r,z_<t), y=D(z_1:T), where x is the input text, r is the emotion reference, z1:Tz_1:T is the generated speech token trajectory, and D is the audio decoder. We intervene during this autoregressive stage, before waveform decoding. Matched synthesis pairs. We next define the neutral–emotion contrast that supplies source and target trajectories. Each contrast uses matched neutral and emotional syntheses that share the same text and speaker setting but differ in the emotion reference. The neutral trajectory uses a neutral reference r(n)r^(n), while the emotional trajectory uses a target-emotion reference r(e)r^(e): τ(n) τ^(n) =(x,r(n),z(n)1:T,y(n)), =(x,r^(n),z^(n)_1:T,y^(n)), τ(e) τ^(e) =(x,r(e),z(e)1:T,y(e)). =(x,r^(e),z^(e)_1:T,y^(e)). Here z1:Tz_1:T is the generated speech token trajectory and y is the decoded waveform for the same text x. Because x and the speaker setting are fixed within each pair, differences between τ(n)τ^(n) and τ(e)τ^(e) define the reference emotion contrast while controlling for lexical content and speaker identity. We use τsrcτ^src for the trajectory that supplies an activation and τtgtτ^tgt for the trajectory being modified. Activation patching. Given a source and target trajectory, activation patching tests whether copying one internal value from the source trajectory into the target trajectory moves the target toward the source condition. For a component site q=(ℓ,t,c)q=( ,t,c), let aq(τ)a_q(τ) denote the activation of component c at layer ℓ and position t on trajectory τ. A patch replaces the target activation with its source counterpart, aq(τtgt)←aq(τsrc),a_q(τ^tgt)← a_q(τ^src), and we score the resulting patched trajectory (Vig et al., 2020; Meng et al., 2022; Heimersheim and Nanda, 2024). The next section defines the scores used for this measurement. Patchable states. Finally, we specify the internal values that can serve as patch sites in the autoregressive model. A patch can replace a residual state, an attention-head output, or an MLP output. At sequence position t, let hℓ,th_ ,t be the residual stream entering layer ℓ . A Transformer block updates this state in two steps, h~ℓ,t=hℓ,t+uℓ,t,hℓ+1,t=h~ℓ,t+mℓ,t, h_ ,t=h_ ,t+u_ ,t, h_ +1,t= h_ ,t+m_ ,t, where uℓ,tu_ ,t is the attention update and mℓ,tm_ ,t is the MLP update. For head-level patches, we decompose the attention update before the output projection, uℓ,t=WℓO[oℓ,1,t‖⋯‖oℓ,H,t],u_ ,t=W^O_ \,[o_ ,1,t\|·s\|o_ ,H,t], where WℓOW^O_ is the layer-ℓ attention output projection, ∥\| denotes concatenation, H is the number of attention heads, and oℓ,j,to_ ,j,t is the output of head j. Thus c can denote a residual stream, head output, MLP output, or block state. Prefix patches use prefix positions; generation-time component patches use generated speech-token positions. 4 Measuring and Localizing Causal Emotion Effects To localize causal emotion effects, activation-patching interventions must be scored on matched neutral and emotional trajectories. We therefore define two complementary metrics for this setting: a codec trajectory score and a late residual direction score. Together, they score patched trajectories and trace reference emotion within the TTS computation; Table 1 reports their stability. 4.1 Metrics for emotion effects Before localizing an emotion effect, we need a scalar score for how much a patched trajectory moves along the neutral-to-emotion contrast. For each matched synthesis pair i, let τi(n) _i^(n) and τi(e) _i^(e) denote the neutral and emotional trajectories defined above, with τi(e) _i^(e) corresponding to the currently scored target emotion k. The codec trajectory score represents a trajectory by the mean embedding of its predicted codec tokens, μ(τ)μ(τ). For target emotion k, let Δkμ=∑i(μ(τi(e))−μ(τi(n))) _k^μ= _i(μ( _i^(e))-μ( _i^(n))). For any trajectory τ paired with the neutral member of pair i, such as a patched trajectory τipatch _i^patch, we define d^kμ d_k^μ =Δkμ/‖Δkμ‖2, = _k^μ/\| _k^μ\|_2, Ck(τ) C_k(τ) =⟨μ(τ)−μ(τi(n)),d^kμ⟩. = μ(τ)-μ( _i^(n)), d_k^μ . Here CkC_k is the codec trajectory score, evaluated relative to the neutral member of the same matched pair. The late residual direction score uses h¯L(τ) h_L(τ), the mean final-layer residual state over generated token positions, where L denotes the final layer index. We define Δkh _k^h =∑i(h¯L(τi(e))−h¯L(τi(n))), = _i( h_L( _i^(e))- h_L( _i^(n))), d^kh d_k^h =Δkh/‖Δkh‖2, = _k^h/\| _k^h\|_2, Rk(τ) R_k(τ) =⟨h¯L(τ)−h¯L(τi(n)),d^kh⟩. = h_L(τ)- h_L( _i^(n)), d_k^h . We use Rk(τ)R_k(τ) as the internal score for target emotion k, with the same pair-relative convention. When the matched pair is clear, we suppress the pair index and write Ck(τ)C_k(τ) or Rk(τ)R_k(τ). Emotion Codec pair Codec split Resid. cent. Anger 0.941 0.998 0.926 Happiness 0.764 0.988 0.903 Sadness 0.903 0.996 0.874 Table 1: Stability of the two emotion metrics over 50 matched pairs. For the codec trajectory score, Codec pair reports the mean pairwise cosine among paired codec-token embedding shifts, and Codec split reports split-half agreement of the global codec direction. Resid. cent. reports the mean cosine to the final-layer residual centroid used by the late residual direction score. Figure 2: Conditioning-prefix source localization and direct source tests. (a) Layer-by-position scans show that emotional-prefix patching into neutral runs has the strongest final-layer residual recovery in the conditioning prefix, with a narrow peak near positions [12,16)[12,16). (b) Direct interventions on the layer-0 [0,32)[0,32) source window validate this source across anger, happiness, and sadness: emotional-prefix patching recovers the emotion effect, while neutral-prefix patching removes it. Ratios near one indicate recovery or removal of nearly the full clean neutral–emotion shift under both codec trajectory and late residual direction scores. Both metrics are stable enough for intervention scoring (Table 1). The codec trajectory score is less consistent for happiness, consistent with more heterogeneous acoustic realizations of happiness (Appendix B.1). The late residual direction score is coherent across emotions and serves as the internal readout target for causal tracing. We use the final layer because it is closest to generated-token prediction, although emotion directions are already separable in earlier layers due to the conditioning prefix (Appendix B.2). These metrics do not establish causality by themselves; they define the targets used to score the interventions below. 4.2 Conditioning-prefix patching for source localization With these metrics, source localization focuses on a natural candidate region in IndexTTS2. The emotion reference is converted into an emotion vector and placed in the GPT input prefix, before the text embeddings and before generated speech tokens are produced. We therefore search the conditioning prefix rather than all generated positions. We test this with prefix patching. Let τ(n)τ^(n) and τ(e)τ^(e) denote paired neutral and emotional trajectories with the same speaker and text but different emotion references. For a conditioning-prefix source site ssrc=(ℓ,)s_src=( ,W), where W is a prefix window at layer ℓ , we run the neutral trajectory while replacing its prefix activations with the emotional counterparts, hℓ,(n)←hℓ,(e),h_ ,W^(n)← h_ ,W^(e), and then measure whether the patched trajectory moves toward the emotional trajectory. We score this movement with either metric defined above, Mk∈Ck,RkM_k∈\C_k,R_k\. Recovery is normalized by the clean gap between the matched emotional and neutral trajectories: Essrc,k=Mk(τssrcpatch)−Mk(τ(n))Mk(τ(e))−Mk(τ(n)).E_s_src,k= M_k( _s_src^patch)-M_k(τ^(n))M_k(τ^(e))-M_k(τ^(n)). Here ssrc=(ℓ,)s_src=( ,W) is the patched prefix site, τssrcpatch _s_src^patch is the neutral trajectory after replacing site ssrcs_src with its emotional counterpart, and MkM_k is either the codec trajectory score or the late residual direction score for target emotion k, using the pair-relative convention defined above. The ratio is defined only when the magnitude of the clean gap exceeds the numerical tolerance of 10−810^-8; otherwise, it is excluded. A value near one means that the patched prefix activations recover most of the clean emotional shift under that metric. For necessity tests, we reverse the patch direction and measure how much of the clean shift is removed when the emotional trajectory receives the neutral prefix. We use these normalized recovery and removal scores to compare prefix sites. 4.3 Localizing and validating the conditioning-prefix source In the IndexTTS2 autoregressive module, the prefix scan localizes the strongest reference-emotion source to the conditioning prefix. Across anger, happiness, and sadness, recovery concentrates near prefix positions [12,16)[12,16), with the strongest signal in low-to-middle patch layers and weaker spread across nearby prefix positions (Figure 2a). Direct tests on this prefix hotspot [12,16)[12,16) confirm that it is a strong partial source: across emotions, [12,16)[12,16) recovers or removes roughly half to two thirds of the clean shift (Appendix F). Rather than using this narrow peak as the only source, we use the full layer-0 conditioning-prefix block [0,32)[0,32) as a conservative layer-0 [0,32)[0,32) source window for subsequent tracing. We validate this source choice with two direct interventions (Figure 2b): patching the emotional prefix into the neutral run should recover the emotional effect, and patching the neutral prefix into the emotional run should remove it. Across eight matched examples per emotion, both interventions are nearly complete under both metrics. The mean codec recovery and removal ratios are 1.00 for all three emotions, and the late residual direction ratios are also approximately 1.00. Thus, for this model, the full conditioning-prefix block provides a conservative layer-0 [0,32)[0,32) source window, not a claim that every token in the window is individually necessary. These controls isolate the reference emotion from simpler confounds. Each matched pair keeps the text and speaker fixed, so the patched contrast is not explained by lexical content or speaker identity. The same source window is used for every validation prompt, so the result does not rely on prompt-specific window choices. This gives a fixed source for the component analyses below. 5 Selecting Candidate Emotion-Carrying Components The localized conditioning-prefix source gives us a place to start, but not a component-level pathway. We next use gradient attribution as a screening step to identify heads and MLPs whose source-induced changes point toward the late emotion readout. This yields a sparse candidate set for each emotion. 5.1 Gradient attribution selects components Directly intervening on every head and MLP would make candidate discovery expensive and hard to separate from validation. We therefore use gradient attribution as a first-order screen, following attribution patching methods for scalable component localization (Nanda, 2023; Kramár et al., 2024). Let τssrcpatch _s_src^patch denote the neutral trajectory after patching the validated conditioning-prefix source ssrc=(0,src)s_src=(0,W_src), with src=[0,32)W_src=[0,32), from the emotional trajectory. For a component c, pilot pair i, and autoregressive prediction step t, define the source-induced activation change Δai,c,t=ai,c,t(τi,ssrcpatch)−ai,c,t(τi(n)), a_i,c,t=a_i,c,t( _i,s_src^patch)-a_i,c,t( _i^(n)), and sum its local first-order effects over prediction steps: Ai,c,k=∑t⟨Δai,c,t,∇ai,c,tRk(τi,ssrcpatch)⟩.A_i,c,k= _t a_i,c,t, _a_i,c,tR_k( _i,s_src^patch) . Here RkR_k is the late residual direction score for target emotion k, and ai,c,t(τ)a_i,c,t(τ) denotes the generation-time output of component c for pair i at step t. The gradient is evaluated on the source-patched trajectory. A large positive Ai,c,kA_i,c,k means that the source-induced change in component c locally increases the emotion readout. For each emotion, we average Ai,c,kA_i,c,k over eight pilot pairs and keep components whose mean score exceeds ηg¯kη g_k, where η=0.005η=0.005 and g¯k=18∑i‖h¯L(τi(e))−h¯L(τi(n))‖2 g_k= 18 _i\| h_L( _i^(e))- h_L( _i^(n))\|_2 is the mean clean pairwise final-layer residual shift over the same pilot pairs. We also require positive attribution on at least seven pairs. Eligible components are then ranked by mean absolute attribution, and the top 30% form the frozen candidate set. Emotion Nodes Heads MLPs Late MLPs Shared Anger 24 15 9 6 13 Happiness 23 13 10 6 13 Sadness 27 21 6 5 13 Table 2: Gradient-attribution candidate sets. Nodes are selected attention-head and MLP components in the frozen top-30% candidate set for each emotion. The Late MLPs column counts selected MLPs in layers 18–23. Shared counts components that appear in all three emotion-specific candidate sets. Emotion Components Pilot recover Pilot remove Valid. recover Valid. remove Anger 24 (15/9) 0.812 / 0.815 0.787 / 0.786 0.814 / 0.815 0.784 / 0.787 Happiness 23 (13/10) 0.883 / 0.886 0.864 / 0.865 0.876 / 0.868 0.855 / 0.855 Sadness 27 (21/6) 0.751 / 0.754 0.754 / 0.755 0.736 / 0.739 0.747 / 0.744 Table 3: Pilot and validation results for frozen top-30% component sets. Each component set is selected on eight pilot pairs and evaluated without retuning on eight validation synthesis pairs. Components reports total selected components, with heads/MLPs in parentheses. Recover is the fraction of the full-source readout shift restored by the selected components; remove is the fraction lost when the same components are replaced with neutral activations in the full-source trajectory. Each recover/remove cell reports mean/median. Table 2 shows that the attribution ranking is sparse, late-MLP-heavy, and partly shared across emotions. The 30% budget is also supported by a compression sweep on anger: the top 10% recovers 0.696 of the readout shift and removes 0.661, the top 15% reaches 0.763 and 0.720, and the top 30% reaches 0.815 and 0.798. Increasing the budget to 50% or 90% gives only smaller additional gains (Appendix G). We therefore use the top 30% as a compact candidate set for validation, not as a claim that lower-ranked components have no effect. The exact frozen component identities are reported in Appendix G.1. This gives a concrete source-to-readout candidate pathway for the intervention tests below. 6 Validating the Source-to-Readout Component-Level Circuit The attribution ranking gives candidate heads and MLPs, but direct interventions are needed to test whether they carry the source effect to the emotion readout. We therefore freeze the selected components and evaluate them on validation synthesis cases. We test sufficiency and necessity, trace source-to-component effects against matched controls, and compare the selected sets with matched random component sets. Emotion Recover Remove Selected Random Selected Random Anger 0.814 0.278 0.784 0.242 Happiness 0.876 0.421 0.855 0.463 Sadness 0.736 0.340 0.747 0.395 Table 4: Matched random controls for component-set validation. Selected cells report validation mean ratios for the frozen component sets. Random cells report the mean over 20 matched random sets with the same size and head–MLP mix. The selected set is larger than every matched random set in all three emotions and both intervention directions. 6.1 Frozen component interventions We apply the sufficiency and necessity logic from source localization (Section 4.2) to the frozen component sets selected above (Section 5). For each emotion, the source window, readout metric, component budget, and selected heads and MLPs are fixed before validation. Thus the following interventions test the attribution-selected set directly, rather than continuing the selection procedure. In the sufficiency test, the neutral trajectory receives the selected head and MLP outputs from the trajectory with the emotional source. In the necessity test, the source-patched neutral trajectory receives the matched clean-neutral outputs at the same components. We normalize both scores by the full-source readout gap between the source-patched neutral and clean-neutral trajectories. A ratio near one means that the frozen component set recovers the readout shift in the sufficiency test or removes it in the necessity test. 6.2 Frozen components recover and suppress the readout The frozen component sets pass the main readout-level validation on synthesis cases that were not used for attribution. Table 3 reports two ratios for each emotion: recover is the sufficiency score, and remove is the necessity score. Across emotions, validation recovery/removal ratios remain close to the pilot ratios, and the expected direction holds for every pair in every condition. Together, the two intervention directions show that the selected components both restore the emotion readout when added to a neutral trajectory and suppress it when replaced with clean-neutral outputs in the source-patched neutral trajectory. This supports the readout-level claim that the frozen sparse sets carry a large fraction of the measured emotion shift. 6.3 Testing component specificity The set-level interventions show that the selected components matter, but do not by themselves rule out a simpler explanation: perhaps many components in similar layers would work equally well. We therefore test specificity at the full-set level by sampling random component sets with the same size and head/MLP mix. This control gives a larger separation (Table 4). Across emotions, selected validation recovery ratios are 0.736–0.876, while matched random sets average 0.278–0.421. The same pattern holds for removal: selected sets remove 0.747–0.855 of the readout shift, while matched random sets average 0.242–0.463. In every emotion and intervention type, the selected set exceeds all 20 random sets. As an auxiliary source-conditioned specificity check, selected components exhibit positive selected-minus-control deltas against cleaner matched alternatives in all three emotions and both intervention directions (Appendix H). Together, these controls show that the result is not explained by patching arbitrary components in similar layers. 6.4 Shared and emotion-specific structure The selected sets combine a common backbone with emotion-specific components. The same 13 components occur in all three top sets. As Table 5 shows, these shared components alone recover 79–86% of the full-set sufficiency effect, while the remaining emotion-specific components recover 58–59%. Both subsets have positive effects on every held-out pair. Table 5: Sufficiency recovery of full, shared-only, and emotion-specific component sets. Parentheses give the fraction of the full-set recovery. Emotion Full Shared only Specific only Anger 0.814 0.686 (84%) 0.470 (58%) Happiness 0.876 0.753 (86%) 0.511 (58%) Sadness 0.736 0.584 (79%) 0.433 (59%) Shared sites nevertheless carry emotion-dependent activation content. Replacing their target-emotion activations with activations from another emotion reduces the target readout in 47 of 48 source–target cases. The mean changes are −33.14-33.14, −40.72-40.72, and −51.46-51.46 for anger, happiness, and sadness, respectively; the overall change is −41.77-41.77, with a 95% confidence interval of [−48.27,−35.28][-48.27,-35.28]. Thus the circuit uses shared computational sites while retaining target-specific activation content and components. Table 6: Decoded acoustic effects of the sparse component intervention and a readout-matched residual-direction baseline. Changes are measured from the matched neutral baseline. The second correlate follows the salient target-reference contrast for each emotion. Emotion Circuit Δ 0 (st) Second correlate Circuit change Residual Δ 0 Ratio Anger +2.77+2.77 [2.44,3.12][2.44,3.12] RMS energy +3.35+3.35 dB [2.99,3.73][2.99,3.73] +0.74+0.74 st 0.27 Happiness +3.57+3.57 [3.18,3.98][3.18,3.98] Spectral centroid +133.06+133.06 Hz [113.37,152.10][113.37,152.10] +0.83+0.83 st 0.23 Sadness −1.01-1.01 [−1.38,−0.55][-1.38,-0.55] RMS energy −3.04-3.04 dB [−3.42,−2.67][-3.42,-2.67] −0.17-0.17 st 0.17 7 Decoded-Audio Effects and Baselines The component tests so far are internal; we next ask whether the same fixed interventions affect decoded waveforms. For each target emotion, we compare 24 matched text-speaker cases under a neutral baseline and a component-patched target intervention. We measure mean F0, active-frame RMS energy, and spectral centroid, calibrating each sign by the corresponding target-minus-neutral reference contrast. We report paired bootstrap 95% confidence intervals, paired sign-permutation tests, and directional consistency. Target-prototype Emo-SIM from emotion2vec (Ma et al., 2024) is reported as a secondary measure in Appendix C. 7.1 Paired acoustic effects Table 6 shows consistent waveform-level changes for all three emotions. Anger increases pitch and energy, happiness increases pitch and spectral brightness, and sadness decreases pitch and energy, matching the corresponding reference contrasts. All reported acoustic changes are significant under paired permutation tests: p≤10−4p≤ 10^-4 with 24/24 directional consistency, except sadness F0 with p=3×10−4p=3× 10^-4 and 23/24 consistency. 7.2 Readout-matched residual-direction steering We compare the component intervention with direct steering along the layer-23 residual emotion direction at every generation step. Steering strength is chosen to match the mean component-induced readout shift. The residual baseline reaches equal or larger internal shifts—56.9 versus 51.0 for anger, 62.1 versus 57.5 for happiness, and 35.1 versus 32.4 for sadness—but produces only 17–27% of the component intervention’s F0 change (Table 6). The selected components therefore transfer reference-derived emotion information to decoded acoustics more effectively than a global readout direction. The effect also persists under top-p sampling: over eight pairs per emotion, F0 changes by +2.15+2.15, +1.69+1.69, and −0.95-0.95 semitones for anger, happiness, and sadness. A second-backbone check on LLaSA-1B further finds that a layer-8 anger-prefix patch recovers 0.855 of the speech-token direction effect over 30 prompts. Appendix D provides the full protocols and confidence intervals. An exploratory blinded A/B check also prefers the intervention for target-emotion strength in 24/24 trials, while 21 of 24 naturalness comparisons are ties (Appendix E). 8 Conclusion This work traces reference-conditioned emotion from an input-prefix source through sparse attention-head and MLP sets to late readouts and decoded speech. The selected component-level circuits recover and suppress most of the measured readout effect on held-out cases, combine a shared computational backbone with emotion-specific content, and outperform matched random sets. Their decoded acoustic effects are substantially stronger than those of readout-matched residual-direction steering, linking the localized components to concrete changes in pitch, energy, and spectral brightness. These results establish a compact causal account of reference-emotion routing in LLM-based TTS under matched-reference conditions. Limitations The main circuit analysis uses IndexTTS2, one speaker/timbre setting, English prompts, three emotions, and one fixed reference file per emotion. The identified components may therefore capture properties of these references in addition to emotion structure. The LLaSA-1B and top-p checks extend the causal pattern to a second backbone and stochastic decoding, but broader tests across speakers, languages, references, and architectures remain open. Our circuit is a component-level source-to-readout account rather than a complete enumeration of every internal edge. Finally, the listening check uses one listener; a multi-listener study would provide stronger perceptual evaluation. Ethical Considerations and Potential Risks Emotion-controllable TTS could be misused to misrepresent a speaker’s affect or to create deceptive voice content, especially when combined with unauthorized voice cloning. Our work is a controlled interpretability study and does not release a deployable editing system. Any deployment of emotion-controllable TTS should require speaker consent, clear disclosure or provenance mechanisms, and safeguards against impersonation and deceptive use. Acknowledgments This work is supported in part by the National Science Foundation (NSF) grant IIS-2451436 and the Commonwealth Cyber Initiative grant HC-2Q26-032. References Bickmore and Picard (2005) T. W. Bickmore and R. W. Picard Establishing and maintaining long-term human-computer relationships. ACM Transactions on Computer-Human Interaction 12 (2), p. 293–327. External Links: Document Cited by: §1. Borsos et al. (2023) Z. Borsos, R. Marinier, D. Vincent, E. Kharitonov, O. Pietquin, M. Sharifi, D. Roblek, O. Teboul, D. Grangier, M. Tagliasacchi, and N. Zeghidour AudioLM: a language modeling approach to audio generation. IEEE/ACM Transactions on Audio, Speech, and Language Processing 31, p. 2523–2533. External Links: Document Cited by: §2. Conmy et al. (2023) A. Conmy, A. Mavor-Parker, A. Lynch, S. Heimersheim, and A. Garriga-Alonso Towards automated circuit discovery for mechanistic interpretability. Advances in Neural Information Processing Systems 36, p. 16318–16352. External Links: Document, Link Cited by: §1, §2. Deng et al. (2025) W. Deng, S. Zhou, J. Shu, J. Wang, and L. Wang IndexTTS: an industrial-level controllable and efficient zero-shot text-to-speech system. arXiv preprint arXiv:2502.05512. External Links: Link Cited by: §1, §2, §3. Du et al. (2024) Z. Du, Q. Chen, S. Zhang, K. Hu, H. Lu, Y. Yang, H. Hu, S. Zheng, Y. Gu, Z. Ma, Z. Gao, and Z. Yan CosyVoice: a scalable multilingual zero-shot text-to-speech synthesizer based on supervised semantic tokens. arXiv preprint arXiv:2407.05407. External Links: Link Cited by: §1, §2, §3. Elhage et al. (2021) N. Elhage, N. Nanda, C. Olsson, T. Henighan, N. Joseph, B. Mann, A. Askell, Y. Bai, A. Chen, T. Conerly, N. DasSarma, D. Drain, D. Ganguli, Z. Hatfield-Dodds, D. Hernandez, A. Jones, J. Kernion, L. Lovitt, K. Ndousse, D. Amodei, T. Brown, J. Clark, J. Kaplan, S. McCandlish, and C. Olah A mathematical framework for transformer circuits. Transformer Circuits Thread. External Links: Link Cited by: §1, §2. Glazer et al. (2026) N. Glazer, Y. Segal-Feldman, H. Segev, A. Shamsian, A. Buchnick, G. Hetz, E. Fetaya, J. Keshet, and A. Navon Beyond transcription: mechanistic interpretability in ASR. In Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 40, p. 37407–37416. External Links: Document, Link Cited by: §2. Goldowsky-Dill et al. (2023) N. Goldowsky-Dill, C. MacLeod, L. Sato, and A. Arora Localizing model behavior with path patching. arXiv preprint arXiv:2304.05969. External Links: Link Cited by: §2. Guo et al. (2023) Z. Guo, Y. Leng, Y. Wu, S. Zhao, and X. Tan PromptTTS: controllable text-to-speech with text descriptions. In 2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), p. 1–5. External Links: Document Cited by: §1, §2. Heimersheim and Nanda (2024) S. Heimersheim and N. Nanda How to use and interpret activation patching. arXiv preprint arXiv:2404.15255. External Links: Link Cited by: §1, §2, §3. Juslin and Laukka (2003) P. N. Juslin and P. Laukka Communication of emotions in vocal expression and music performance: different channels, same code?. Psychological Bulletin 129 (5), p. 770–814. External Links: Document Cited by: §1. Kramár et al. (2024) J. Kramár, T. Lieberum, R. Shah, and N. Nanda AtP*: an efficient and scalable method for localizing LLM behaviour to components. arXiv preprint arXiv:2403.00745. External Links: Link Cited by: §2, §5.1. Li et al. (2026) H. Li, Y. Liu, Y. Sun, H. Shi, L. Qu, and T. Li EMORL-TTS: reinforcement learning for fine-grained emotion control in LLM-based TTS. In 2026 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), p. 17272–17276. External Links: Document Cited by: §2. Ma et al. (2024) Z. Ma, Z. Zheng, J. Ye, J. Li, Z. Gao, S. Zhang, and X. Chen Emotion2vec: self-supervised pre-training for speech emotion representation. In Findings of the Association for Computational Linguistics: ACL 2024, Bangkok, Thailand, p. 15747–15760. External Links: Document, Link Cited by: §7. Meng et al. (2022) K. Meng, D. Bau, A. Andonian, and Y. Belinkov Locating and editing factual associations in GPT. In Advances in Neural Information Processing Systems, Vol. 35, p. 17359–17372. External Links: Document, Link Cited by: §1, §2, §3. Nanda (2023) N. Nanda Attribution patching: activation patching at industrial scale. Note: Online articleAccessed: 2026-08-25 External Links: Link Cited by: §5.1. Olah et al. (2020) C. Olah, N. Cammarata, L. Schubert, G. Goh, M. Petrov, and S. Carter Zoom in: an introduction to circuits. Distill. External Links: Document, Link Cited by: §1, §2. Olsson et al. (2022) C. Olsson, N. Elhage, N. Nanda, N. Joseph, N. DasSarma, T. Henighan, B. Mann, A. Askell, Y. Bai, A. Chen, T. Conerly, D. Drain, D. Ganguli, Z. Hatfield-Dodds, D. Hernandez, S. Johnston, A. Jones, J. Kernion, L. Lovitt, K. Ndousse, D. Amodei, T. Brown, J. Clark, J. Kaplan, S. McCandlish, and C. Olah In-context learning and induction heads. Note: Transformer Circuits Thread External Links: Link Cited by: §2. Paek et al. (2025) N. Paek, Y. Zang, Q. Yang, and R. Leistikow Learning interpretable features in audio latent spaces via sparse autoencoders. arXiv preprint arXiv:2510.23802. External Links: Link Cited by: §2. Picard (1997) R. W. Picard Affective computing. MIT Press, Cambridge, MA. External Links: ISBN 978-0-262-16170-1, Link Cited by: §1. Scherer (2003) K. R. Scherer Vocal communication of emotion: a review of research paradigms. Speech Communication 40 (1–2), p. 227–256. External Links: Document Cited by: §1. Vig et al. (2020) J. Vig, S. Gehrmann, Y. Belinkov, S. Qian, D. Nevo, Y. Singer, and S. Shieber Investigating gender bias in language models using causal mediation analysis. In Advances in Neural Information Processing Systems, Vol. 33, p. 12388–12401. External Links: Link Cited by: §1, §2, §3. Wang et al. (2023a) C. Wang, S. Chen, Y. Wu, Z. Zhang, L. Zhou, S. Liu, Z. Chen, Y. Liu, H. Wang, J. Li, L. He, S. Zhao, and F. Wei Neural codec language models are zero-shot text to speech synthesizers. arXiv preprint arXiv:2301.02111. External Links: Link Cited by: §1, §2, §3. Wang et al. (2023b) K. R. Wang, A. Variengien, A. Conmy, B. Shlegeris, and J. Steinhardt Interpretability in the wild: a circuit for indirect object identification in GPT-2 small. In International Conference on Learning Representations, External Links: Link Cited by: §2. Wang et al. (2025) X. Wang, M. Jiang, Z. Ma, Z. Zhang, S. Liu, L. Li, Z. Liang, Q. Zheng, R. Wang, X. Feng, W. Bian, Z. Ye, S. Cheng, R. Yuan, Z. Zhao, X. Zhu, J. Pan, L. Xue, P. Zhu, Y. Chen, Z. Li, X. Chen, L. Xie, Y. Guo, and W. Xue Spark-TTS: an efficient LLM-based text-to-speech model with single-stream decoupled speech tokens. arXiv preprint arXiv:2503.01710. External Links: Link Cited by: §1, §2, §3. Xie et al. (2025) T. Xie, S. Yang, C. Li, D. Yu, and L. Liu EmoSteer-TTS: fine-grained and training-free emotion-controllable text-to-speech via activation steering. arXiv preprint arXiv:2508.03543. External Links: Link Cited by: §2. Yang et al. (2024) D. Yang, S. Liu, R. Huang, C. Weng, and H. Meng InstructTTS: modelling expressive TTS in discrete latent space with natural language style prompt. IEEE/ACM Transactions on Audio, Speech, and Language Processing 32, p. 2913–2925. External Links: Document Cited by: §1, §2. Yang et al. (2025) G. Yang, C. Yang, Q. Chen, Z. Ma, W. Chen, W. Wang, T. Wang, Y. Yang, Z. Niu, W. Liu, F. Yu, Z. Du, Z. Gao, S. Zhang, and X. Chen EmoVoice: LLM-based emotional text-to-speech model with freestyle text prompting. In Proceedings of the 33rd ACM International Conference on Multimedia, p. 10748–10757. External Links: Document Cited by: §1, §2. Zhang et al. (2025) S. Zhang, A. Mehrish, Y. Li, and S. Poria PROEMO: prompt-driven text-to-speech synthesis based on emotion and intensity control. arXiv preprint arXiv:2501.06276. External Links: Link Cited by: §2. Zhou et al. (2026) S. Zhou, Y. Zhou, Y. He, X. Zhou, J. Wang, W. Deng, and J. Shu IndexTTS2: a breakthrough in emotionally expressive and duration-controlled auto-regressive zero-shot text-to-speech. In Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 40, p. 35139–35148. External Links: Document, Link Cited by: §1, §2. Appendix A Experimental Setup Details Model and tracing configuration. All experiments use the released IndexTTS2-style autoregressive GPT/T2S module with the checkpoint configuration loaded from index-tts/checkpoints. The model is run in evaluation mode, and the accelerated inference engine is disabled during tracing so that PyTorch hooks can capture and patch internal states. Unless otherwise noted, generation traces are capped at 200 generated speech token steps. Text and reference manifests. The paired manifests used in our experiments contain 50 English text prompts and four reference conditions per prompt: neutral, anger, happiness, and sadness, for 200 rows in each manifest. They use a single speaker identifier, 0001. Within each neutral–emotion pair, the text and speaker/timbre reference are fixed, while the style reference is an utterance expressing the target emotion. The exact style-reference utterance used for each run is specified in its corresponding manifest. Pair construction. For each text id i and target emotion k∈anger,happy,sadk∈\anger,happy,sad\, we construct a pair identifier such as text0_anger consisting of a neutral trajectory and an emotional trajectory for the same prompt. This yields 150 possible neutral–emotion pairs: 50 text prompts for each of three non-neutral target emotions. Pair construction requires both the neutral row and the target-emotion row to exist in the manifest and requires the corresponding cached trace to be present. Pairs that do not satisfy these conditions are excluded before analysis. Generation and random seeds. The main experiments use deterministic greedy codec-token decoding. We set the Python, NumPy, PyTorch, and CUDA random seeds to 42 in the patching scripts. A separate robustness run uses top-p sampling with temperature 0.9 and p=0.95p=0.95, holding the sampling seed fixed across intervention roles within each matched pair. Activation caches. For each pair, the cached trace stores both neutral and emotional trajectories. The main cache captures residual streams, MLP outputs, attention-head outputs, and full prefill activations for all Transformer layers. Generation-time caches store the activation used to predict each next codec token, along with metadata including prefix length, generated token id, and generation length. Prefix source interventions capture donor prefill residual states from the emotional prompt and patch them into the corresponding neutral prompt at the same layer and prefix positions. Pair regimes and validation split. We use two pair regimes. First, coarse source-localization scans are run over the full 50 pairs per target emotion when available; the merged full-scan outputs record num_pairs_used=50. Second, higher-cost interventions use fixed 8-pair pilot sets for each target emotion. The pilot ids are text0, text1, and text10–text15, paired with the target emotion suffix. Components are selected on these eight pilot prompts and evaluated without retuning on eight validation synthesis cases. The matched random component controls use the same validation synthesis cases and the same sufficiency and necessity metrics. The source-conditioned component specificity control also uses the validation synthesis cases and keeps the frozen component membership fixed. Decoded-audio analyses use 24 matched synthesis pairs per emotion, and the top-p check uses eight pairs per emotion. Appendix B Metric Validation and Readout Diagnostics B.1 Happiness codec-direction heterogeneity Happiness has lower pair-level codec agreement than anger and sadness in Table 1. To check whether this is driven by a few anomalous pairs, we remove the most inconsistent happy pairs and recompute the codec-direction statistics. Table 7 shows that pairwise agreement increases only modestly, while split-half agreement stays near 0.99, suggesting a reproducible but more heterogeneous happiness direction. Removed Pair cos. Split cos. 0 0.764 0.988 1 0.769 0.986 3 0.780 0.988 5 0.789 0.987 10 0.810 0.987 Table 7: Outlier removal for the Happiness codec direction. Removing the least aligned happiness pairs modestly increases pairwise agreement but leaves split-half agreement high. B.2 Residual direction clustering To examine when reference-conditioned emotion becomes separable, we visualize paired residual differences at the first autoregressive prediction step across layers. Figure 3: Emotion clustering at the first autoregressive prediction step. Each point is a paired sample-level emotion direction, computed as the residual difference between an emotional trajectory and its neutral counterpart at the same layer and prediction step. t-SNE is fit independently within each layer. This figure measures layerwise separability at a single prediction step; the whole-generation residual readout used for component selection and validation is computed separately. Figure 3 shows the resulting first-step layerwise separation; the whole-generation residual readout is evaluated separately. Appendix C Decoded-Audio Statistical Details We retain calibrated target-prototype Emo-SIM as an auxiliary decoded-audio measure. Table 8 reports paired bootstrap intervals, win counts, and sign-permutation tests over the same 24 pairs per emotion used in the acoustic analysis. Happiness and sadness show positive calibrated changes; the anger interval includes zero. The primary decoded-audio evidence in Section 7 is therefore the more direct and consistent acoustic analysis. Table 8: Paired statistics for calibrated Emo-SIM change. Emotion Mean [95% CI] Wins p Anger +0.323+0.323 [−0.209,0.833][-0.209,0.833] 17/24 0.243 Happiness +0.730+0.730 [0.438,1.047][0.438,1.047] 20/24 2×10−52× 10^-5 Sadness +0.396+0.396 [0.105,0.731][0.105,0.731] 19/24 0.0156 Appendix D Decoding and Second-Backbone Robustness Top-p decoding. We rerun the frozen component interventions with temperature 0.9 and top-p=0.95p=0.95 on eight pairs per emotion. The sampling seed is shared across roles within each pair. The resulting F0 changes preserve the greedy-decoding directions (Table 9), and the target readout moves in the intended direction in all 24 cases. Table 9: F0 effects under top-p decoding. Intervals are paired bootstrap 95% confidence intervals over eight pairs per emotion. Emotion Δ 0 in semitones [95% CI] Anger +2.15+2.15 [1.29,3.00][1.29,3.00] Happiness +1.69+1.69 [1.20,2.22][1.20,2.22] Sadness −0.95-0.95 [−1.46,−0.45][-1.46,-0.45] LLaSA-1B source localization. Component identities are architecture-specific, so the second-backbone test repeats the source-sufficiency question rather than transferring IndexTTS2 components. On 30 matched anger–neutral prompts, a right-aligned layer-8 LLaSA prefix patch recovers 0.855 of the generated speech-token anger-direction effect (95% CI [0.771,0.931][0.771,0.931]; 30/30 directional consistency). A neutral random-window control is near zero at −0.044-0.044 (95% CI [−0.148,0.043][-0.148,0.043]). Appendix E Exploratory Listening Check One of the authors completed a blinded A/B comparison over 24 matched trials, with eight trials per target emotion. This author-conducted evaluation is intended only as an exploratory sanity check; no external participants were recruited, and the results should not be interpreted as population-level perceptual evidence. Each trial compared the neutral baseline with the sparse component intervention using the same text and reference setup. The intervention was preferred for target-emotion strength in all 24 trials; naturalness was judged as a tie in 21 trials (Table 10). Table 10: Single-listener blinded A/B results. Naturalness cells report intervention/baseline/tie counts. The reported p-values are item-level one-sided exact sign tests. Emotion Emotion pref. p Naturalness Anger 8/8 0.0039 0/1/7 Happiness 8/8 0.0039 1/1/6 Sadness 8/8 0.0039 0/0/8 Overall 24/24 5.96×10−85.96× 10^-8 1/2/21 Appendix F Conditioning-Prefix Source Localization Details The conditioning-prefix source scan evaluates eight non-overlapping four-token prefix windows from [0,4)[0,4) to [28,32)[28,32). The narrow hotspot [12,16)[12,16) is selected from these scans. The broader layer-0 [0,32)[0,32) source window is then used for downstream validation. Source-window validation records the neutral and emotional prefix lengths and skips a pair if the selected source window exceeds either prefix length. In the current validation set, the [0,32)[0,32) envelope fits all reported pairs. Table 11 reports the corresponding narrow-window sufficiency and necessity results. Emotion Codec suff. Resid. suff. Codec nec. Resid. nec. Anger 0.62 0.60 0.47 0.54 Happiness 0.55 0.55 0.73 0.65 Sadness 0.59 0.53 0.59 0.64 Table 11: Direct tests on the narrow prefix hotspot [12,16)[12,16). Ratios are averaged over 50 matched pairs per emotion. The hotspot recovers or removes a substantial fraction of the clean neutral–emotion shift, but does not fully account for the effect; the main tracing experiments therefore use the broader layer-0 [0,32)[0,32) conditioning-prefix block as the source window. Appendix G Component Selection and Set-Level Validation Details Component-set selection and validation boundary. Component candidates are selected from a source-conditioned attribution graph using the fixed layer-0 [0,32)[0,32) source window and a fixed layer-23 residual emotion readout. Canonical top-30% sets are frozen before the direct sufficiency and necessity tests. The resulting sets contain 24 nodes for anger, 27 nodes for sadness, and 23 nodes for the happiness configuration used in the cross-emotion component-set summary. These validation tests therefore evaluate fixed component sets rather than retuning membership during the intervention trajectory. G.1 Canonical Component Identities The evaluated component universe contains 480 attention heads and 24 MLPs across 24 Transformer layers, for 504 components in total. Table 12 lists the three frozen sets in mean-absolute-attribution rank order. We write LxHyLxHy for attention head y in layer x, and LxMLPLxMLP for the MLP in layer x. The sets aggregate attribution across the eight pilot pairs and are not reselected on validation pairs; attribution magnitudes and within-set ranks may vary across individual pairs. Component identities are specific to the studied checkpoint and reference configuration. Table 12: Frozen canonical component identities. Each column follows its emotion-specific attribution ranking. Bold entries are the 13 components shared by all three sets. Anger Happiness Sadness L21MLP L22MLP L0H2 L22MLP L21MLP L21MLP L20MLP L0H2 L22MLP L18MLP L20MLP L20MLP L16H8 L19MLP L0MLP L17MLP L6H2 L0H16 L19MLP L0MLP L6H2 L6H2 L16H8 L16H8 L9H14 L23MLP L9H19 L15H19 L17MLP L19MLP L12H19 L18MLP L12H0 L23MLP L9H14 L15H19 L16MLP L0H16 L23MLP L12H0 L14H13 L11H9 L8H12 L10H19 L12H9 L14H13 L12H19 L9H12 L15MLP L12H9 L13H13 L14H19 L11H10 L12H19 L11H10 L15H19 L10H19 L10H19 L10H4 L10H12 L14H18 L5MLP L11H10 L10H12 L6MLP L11H14 L13H13 L14H18 L9H14 L10H4 L10H16 L14H4 L8H14 L14H13 Figure 4: Source-conditioned component specificity controls. Each selected component is compared with five cleaner matched alternatives: same-layer non-selected heads for head components, and nearby same-type non-selected MLPs for MLP components. Bars report pair-level selected-minus-control deltas after averaging over selected components within each validation pair; error bars are bootstrap confidence intervals over the eight validation pairs. Positive values indicate that selected components carry more source-conditioned readout effect than matched alternatives. Component-budget sweep. We use an auxiliary anger budget sweep to test how quickly the source-to-readout effect appears as components are added. Retaining only the top 10% of nodes gives a sufficiency recovery ratio of 0.696 and a necessity removal ratio of 0.661. Increasing the budget to the top 15% raises these values to 0.763 and 0.720, and the top 30% reaches 0.815 and 0.798. Larger budgets add only modest gains: the top 50% gives 0.831 recovery and 0.828 removal, while the top 90% gives 0.870 recovery and 0.862 removal. This compression curve supports a sparse component-set interpretation. The effect is not confined to a single component, but much of the measurable source-to-readout effect is captured by a relatively small subset of ranked nodes. Group blockout analyses further indicate that late MLPs contribute a substantial fraction of the effect: removing L20L20–L23L23 MLPs from the anger top-30% component set reduces the necessity removal ratio from 0.798 to 0.677 and the sufficiency recovery ratio from 0.815 to 0.698. Removing middle-layer heads also reduces the effect, especially the head-specific contribution, but the late-MLP removal produces the larger overall decrease in the full component-set score. Appendix H Source-Conditioned Component Specificity Controls We run an auxiliary component-level control to test whether selected components outperform cleaner matched alternatives after conditioning-prefix source patching. The layer-0 [0,32)[0,32) source window remains fixed, and the readout remains the whole-generation layer-23 residual emotion readout. For each selected head, we sample five same-layer heads that are not in the frozen selected set. For each selected MLP, we sample five same-type MLP controls, preferring nearby layers within ±2± 2, expanding to ±4± 4, and then filling from the closest non-selected MLPs if needed. The control manifest records the candidate-pool size, matching rule, and selected controls for every component. For each validation pair and selected component, we compare the selected component score with the mean score of its five controls. We then average these selected-minus-control deltas over selected components within the same validation pair and bootstrap over the eight pairs. This keeps the validation pair, rather than the component row, as the statistical unit. The deltas are positive for all three emotions and both intervention directions: recovery ranges from +0.038+0.038 to +0.064+0.064, and removal ranges from +0.038+0.038 to +0.074+0.074, with bootstrap confidence intervals excluding zero in every condition (Figure 4). Appendix I Cross-Emotion Component Swaps The canonical component rankings share a common backbone, but the activations at these sites are emotion dependent. For each ordered source–target emotion pair, we replace the target-emotion activations at shared sites with donor activations from the other emotion. Table 13 reports the target readout change relative to the target-source intervention. Cross-emotion donors reduce the target readout in 47 of 48 cases, with an overall mean change of −41.77-41.77 and a 95% confidence interval excluding zero. Table 13: Cross-emotion shared-component swaps. Negative values mean that another emotion’s donor activations weaken the target readout. Target Mean change 95% CI Lower Anger −33.14-33.14 [−47.44,−18.84][-47.44,-18.84] 15/16 Happiness −40.72-40.72 [−52.73,−28.70][-52.73,-28.70] 16/16 Sadness −51.46-51.46 [−53.47,−49.46][-53.47,-49.46] 16/16 Overall −41.77-41.77 [−48.27,−35.28][-48.27,-35.28] 47/48