Paper deep dive
Do SpeechLMs Hear Their Own Opinions? Diagnosing and Mitigating Previous-Belief Contamination in Streaming Emotion Understanding
Haoyue Liu, Zhichao Wang, Ye Chen, Haonan Deng, Xiaoying Tang
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 96%
Last extracted: 8/24/2026, 5:16:21 AM
Summary
This paper identifies and diagnoses 'Previous-Belief Contamination' (PBC), a failure mode in Streaming Speech Language Models (SpeechLMs) where the model's own previous prediction distorts its perception of current audio. The authors propose EmoUpdate, a training-free framework that separates current audio perception from historical state revision using an acoustic firewall, a causal belief filter, and a decontamination operator, significantly improving accuracy on streaming emotion benchmarks.
Entities (8)
Relation Signals (6)
Acoustic History Firewall → partof → EmoUpdate
confidence 99% · EmoUpdate... through three components: (1) a prior-blind acoustic firewall...
Causal Belief Filter → partof → EmoUpdate
confidence 99% · EmoUpdate... through three components: (2) an evidence-shrunk causal belief filter...
Decontamination Operator → partof → EmoUpdate
confidence 99% · EmoUpdate... through three components: (3) a closed-form decontamination operator...
EmoUpdate → mitigates → Previous-Belief Contamination
confidence 98% · To address PBC, we introduce EmoUpdate... EmoUpdate achieves the best step accuracy... over the strongest controlled baselines.
Previous-Belief Contamination → affects → Streaming Emotion Understanding
confidence 95% · We show that this history conditioning can distort current perception... revealing a failure we call previous-belief contamination (PBC).
Qwen2-Audio → evaluatedon → CREMA-D-Stream
confidence 95% · On a balanced CREMA-D-Stream counterfactual diagnostic... (CREMA-D-Stream, Qwen2-Audio)
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Streaming emotion understanding uses historical state while continuously interpreting current audio, often feeding the model's previous prediction back as context. We show that this history conditioning can distort current perception. On a balanced CREMA-D-Stream counterfactual diagnostic, changing only the injected previous emotion label while holding the audio fixed reduces current-audio accuracy from 72.50% to 30.42% and flips 65.69% of predictions. The effect is strongly label-asymmetric, with prior pull ranging from 4.76% to 98.20%, revealing a failure we call previous-belief contamination (PBC). To address PBC, we introduce EmoUpdate, a training-free framework that separates current-audio perception from historical state revision through three components: (1) a prior-blind acoustic firewall that prevents historical state from entering perception; (2) an evidence-shrunk causal belief filter that introduces history only after observation formation and retains label-asymmetric transition structure only when supported by observed evidence; and (3) a closed-form decontamination operator derived from the same counterfactual measurements for serving stacks where firewalling is unavailable. Across four SpeechLMs and two streaming emotion benchmarks, EmoUpdate achieves the best step accuracy and state-balanced accuracy in all eight model--benchmark settings, improving S-BAcc by up to 69.71 points and step accuracy by up to 38.41 points over the strongest controlled baselines.
Tags
Links
- Source: https://arxiv.org/abs/2608.20769v1
- Canonical: https://arxiv.org/abs/2608.20769v1
Trouble viewing inline? Open PDF directly →
Full Text
92,732 characters extracted from source content.
Expand or collapse full text
Do SpeechLMs Hear Their Own Opinions? Diagnosing and Mitigating Previous-Belief Contamination in Streaming Emotion Understanding Haoyue Liu Zhichao Wang Ye Chen Affiliation: XJTU-POLIMI Joint School, Xi’an Jiaotong University, Xi’an 710049, China Haonan Deng Affiliation: University of California, Berkeley Xiaoying Tang Affiliation: Shenzhen Future Network of Intelligence Institute (FNii-Shenzhen) [0.6em] School of Science Engineering The Chinese University of Hong Kong Shenzhen 518172 China Abstract Streaming emotion understanding uses historical state while continuously interpreting current audio, often feeding the model’s previous prediction back as context. We show that this history conditioning can distort current perception. On a balanced CREMA-D-Stream counterfactual diagnostic, changing only the injected previous emotion label while holding the audio fixed reduces current-audio accuracy from 72.50% to 30.42% and flips 65.69% of predictions. The effect is strongly label-asymmetric, with prior pull ranging from 4.76% to 98.20%, revealing a failure we call previous-belief contamination (PBC). To address PBC, we introduce EmoUpdate, a training-free framework that separates current-audio perception from historical state revision through three components: (1) a prior-blind acoustic firewall that prevents historical state from entering perception; (2) an evidence-shrunk causal belief filter that introduces history only after observation formation and retains label-asymmetric transition structure only when supported by observed evidence; and (3) a closed-form decontamination operator derived from the same counterfactual measurements for serving stacks where firewalling is unavailable. Across four SpeechLMs and two streaming emotion benchmarks, EmoUpdate achieves the best step accuracy and state-balanced accuracy in all eight model–benchmark settings, improving S-BAcc by up to 69.71 points and step accuracy by up to 38.41 points over the strongest controlled baselines. 1 Introduction Audio-language models can reason directly over speech without first reducing it to a transcript (Chu et al. 2024; Xu et al. 2025; Abouelenin et al. 2025), making them well suited to streaming emotion understanding, where a system tracks a speaker’s emotional state as audio continuously arrives (Guo et al. 2026; Wang et al. 2026a; Song et al. 2026). Such tasks require a model to interpret the current audio while also using historical state, so its previous prediction is fed back as historical context. We formalize this design below as direct history conditioning. We ask how this history conditioning affects current perception. Once the previous state enters the perception prompt, it is no longer only a summary of the past. It becomes textual context that directly interacts with the current acoustic evidence. The model may copy the previous judgment, preserve it through a genuine emotional shift, or more subtly reshape the current posterior. The system, in short, begins to hear its own opinions. This failure can be mistaken for ordinary recognition error, but it is structurally different. The current audio remains unchanged, yet the resulting observation changes when only the injected historical label is replaced. It also differs from context-induced biases caused by external textual evidence (Goyal et al. 2024). Here, the contaminating variable is self-generated and recurrent, so an erroneous state can re-enter subsequent inference and propagate over time. To make this failure measurable, we introduce a counterfactual intervention protocol that holds the current audio, instruction, and decoding fixed while varying only the injected previous emotion label and measuring the resulting posterior change. On a balanced CREMA-D-Stream diagnostic, this intervention reveals a failure we call previous-belief contamination (PBC): history exposure reduces current-audio accuracy by 42.08 points, flips 65.69% of predictions, and pulls 71.00% of wrong-prior conditions toward the injected label. More importantly, the effect is strongly label-asymmetric, with prior pull ranging from 4.76% to 98.20% across emotion labels. A single uniform stickiness or global calibration parameter is therefore insufficient to characterize the contamination. Prompt wording alone is also insufficient: seven of ten grounding candidates explicitly instruct the model to ignore the previous label, yet wrong-prior pull never falls below 59.7%. These results suggest that the problem is not simply how strongly history is weighted. Historical state participates in forming the current observation itself, which explains both the degradation under direct history conditioning and the gains from separating perception from state revision shown in Figure 1. Figure 1: Previous-belief contamination and its repair. (a) A wrong injected prior collapses current-audio accuracy (CREMA-D-Stream, Qwen2-Audio). (b) EmoUpdate attains the best causal step accuracy in all eight combinations of models and benchmarks. We therefore introduce EmoUpdate (Figure 2), a training-free streaming inference framework. Its key idea is not to further adjust the weight of historical information, but to change where history enters the system. The current audio first forms an independent acoustic observation, and historical state is introduced only afterward for revision. Specifically, EmoUpdate (1) isolates historical state during perception so that current acoustic evidence is not directly shaped by the previous prediction; (2) performs post-perception causal revision with an evidence-shrunk transition rule that preserves label-asymmetric dynamics only when supported by transition evidence and otherwise reduces to a symmetric prior; and (3) converts the same counterfactual measurements into a closed-form decontamination repair when such isolation is unavailable at the serving interface. Historical information therefore no longer participates in extracting current acoustic evidence and instead acts only during subsequent state revision. We summarize our contributions as follows: • We identify and quantify previous-belief contamination (PBC), a self-referential failure mode in streaming SpeechLMs. Controlled interventions show that a model’s own previous prediction can alter its perception of identical audio, with strong label asymmetry. • We introduce EmoUpdate, a streaming inference paradigm that separates current perception from historical state revision. It first forms a prior-blind acoustic observation, then performs evidence-shrunk causal revision that retains label-asymmetric transitions only when supported by evidence. When firewalling is unavailable, the same measurements yield a closed-form decontamination repair without training or extra deployment calls. • We systematically validate the effectiveness and mechanism of EmoUpdate. Across four SpeechLMs and two streaming emotion benchmarks, EmoUpdate achieves the best step accuracy and state-balanced accuracy in all eight settings, with gains of up to 38.41 and 69.71 percentage points over the strongest controlled baselines. 2 Related Work Speech emotion understanding and streaming emotion modeling. Large audio-language models increasingly support direct speech input and instruction following (Chu et al. 2024; Xu et al. 2025; Abouelenin et al. 2025; Tang et al. 2024; Gong et al. 2024). Recent work studies speech emotion through prosodic prompting and lexical–paralinguistic conflict (Wang et al. 2026b; Pang et al. 2026), while HumDial-EIBench and EmoS evaluate multi-turn and streaming emotional trajectories (Wang et al. 2026a; Guo et al. 2026). More broadly, conversational emotion recognition models temporal dependence through emotional inertia, shift-aware context, speaker-conditioned priors, and learned update controllers (Liu et al. 2024b; Zha et al. 2025; Kaplan & Wermter 2026; Shen et al. 2024; Song et al. 2026). These approaches primarily model how emotion evolves over time. We study a different question: whether the system’s own previous prediction changes its perception of identical current speech. We diagnose this effect through direct intervention on the previous state and separate current-audio perception from subsequent historical state revision. Self-conditioning, prompt bias, and error accumulation. Conditioning a model on its own previous outputs is a classical source of error accumulation (Bengio et al. 2015; Ranzato et al. 2015; Lamb et al. 2016; Arora et al. 2022), while prompted language models can also be influenced by asserted beliefs, prompt position, or conflicting context (Sharma et al. 2024; Xu et al. 2024; Huang et al. 2024; Zhao et al. 2021; Liu et al. 2024a; Goyal et al. 2024). We focus on a different closed-loop setting in which the model’s own previous output re-enters the next perception prompt. PBC differs from ordinary exposure bias because the previous self-generated label does not merely affect a later prediction; it changes the model’s observation of unchanged current audio itself. We expose this failure through controlled counterfactual intervention and separate current-audio grounding from subsequent historical state revision. Appendix A expands this comparison and further covers speech representations and emotion corpora. 3 Method We formulate streaming emotion understanding as causal state estimation, then describe the intervention used to diagnose PBC and the three components of EmoUpdate: the acoustic history firewall, the evidence-shrunk causal belief filter, and the decontamination fallback for stacks that cannot be firewalled (Figure 2). Figure 2: From entangled perception to causal belief revision. (1) Direct history conditioning writes the previous label into the perception prompt; a counterfactual intervention on that label quantifies previous-belief contamination. (2) EmoUpdate restores the observation with a prior-blind acoustic firewall, then revises it with an evidence-shrunk causal filter. (3) Trajectory quality improves; a prompt-free decontamination operator is the fallback where firewalling is unavailable. 3.1 Streaming Emotion Belief Revision An episode is a sequence ℰ=(xt,yt)t=1TE=\(x_t,y_t)\_t=1^T of T chunks, where xtx_t is the current speech chunk, Y is the fixed emotion ontology, and yt∈y_t is the state of that chunk. After receiving xtx_t, a causal system outputs a belief distribution ptp_t and label y^t=argmaxypt(y) y_t= _yp_t(y) without observing future audio. The update action is keep when y^t=y^t−1 y_t= y_t-1 and revise otherwise. This exposes two opposing errors: inertia, retaining a previously correct label after a true shift, and volatility, destroying a correct belief while the gold state remains stable. A direct history-conditioned SpeechLM computes qt(b)=fθ(xt,s,bt−1),q_t^(b)=f_θ(x_t,s,b_t-1), (1) where qt(b)q_t^(b) is the resulting label posterior over Y, fθf_θ is the frozen SpeechLM, s is an instruction, and bt−1∈b_t-1 is the previous emitted state label, rendered into the prompt as text. In deployment, bt−1=y^t−1b_t-1= y_t-1. Equation 1 therefore uses bt−1b_t-1 both as temporal evidence and as text inside the mechanism that extracts current evidence. Our central claim is that these roles must be separated. 3.2 Counterfactual Previous-Belief Intervention For a fixed current chunk xtx_t, we first obtain the control posterior qt(∅)=fθ(xt,s)q_t^( )=f_θ(x_t,s), then insert each candidate previous label b∈ℬ⊆b while keeping all other inputs fixed: qt(b)=fθ(xt,s,b),b∈ℬ.q_t^(b)=f_θ(x_t,s,b), b . (2) We quantify PBC with three complementary measures. Prior flip rate counts interventions for which argmaxqt(b)≠argmaxqt(∅) q_t^(b)≠ q_t^( ). Wrong-prior pull is the fraction of conditions with b≠ytb≠ y_t for which argmaxqt(b)=b q_t^(b)=b. Finally, posterior drift averages the Jensen–Shannon divergence DJS(qt(b)∥qt(∅))D_JS(q_t^(b)\|q_t^( )). The test is diagnostic: its multiple calls are never used during deployment. 3.3 Acoustic History Firewall EmoUpdate constrains the perception input: the call may not receive the previous label, filtered state, episode history, future audio, speaker identifier, or filename cue. Each call is stateless: no conversation history, no carried system prompt, and no reused decoding context. Every prompt actually sent is recorded by SHA-256 in the released audit logs, so prior-blindness is checkable rather than asserted. It computes qt=fθ(xt,s⋆),q_t=f_θ(x_t,s ), (3) using only the current audio and the selected acoustic instruction s⋆s , which asks for pitch, energy, rhythm, pace, vocal tension, and voice quality, and for exactly one ontology label. The posterior is read densely from label-token probabilities, not from self-reported confidence. The firewall is an input constraint, not a particular prompt string. We instantiate it with a fixed ten-candidate development pool whose members use compact alternative strategies for eliciting current-audio evidence while preserving the same prior-blind input contract. Selection uses development data only and is locked before evaluation under a pre-specified trajectory-metric ranking with a final-accuracy floor (Appendix E states the exact rule and lists every candidate). This is black-box selection over readable instructions, not SFT, RL, or soft-prompt learning. 3.4 Causal Belief Filter The prior-blind posterior is fused with history only after perception. Here causal filtering denotes strictly forward online inference using only current and past observations; the effect of prior exposure is established separately by the controlled intervention of Section 3.2. Let ρ∈[0,1]ρ∈[0,1] denote the probability that the emotion remains unchanged and K=||K=|Y|. The one-step predictive prior is p¯t(y)=ρpt−1(y)+1−ρK−1(1−pt−1(y)). p_t(y)=ρ p_t-1(y)+ 1-ρK-1 (1-p_t-1(y) ). (4) We then apply a causal Bayesian update, pt(y)=qt(y)p¯t(y)∑y′∈qt(y′)p¯t(y′).p_t(y)= q_t(y) p_t(y) _y q_t(y ) p_t(y ). (5) This is the filtering recursion of a symmetric first-order hidden Markov model (Rabiner 1989), with p1=q1p_1=q_1 and ρ selected from five pre-specified development values; the filtered argmax induces an explicit keep/revise action. This symmetric filter is the base case extended by the evidence-shrunk transition prior in the next subsection. The decontamination operator instead provides a separate fallback when the perception interface cannot be firewalled. 3.5 Evidence-Shrunk Transition Prior The uniform off-diagonal term treats all revisions as exchangeable. We therefore let development transition statistics encode label-asymmetric destinations, shrunk toward the uniform null when evidence is weak. For emotion states i,j∈i,j , the off-diagonal transition is A(i→j)=(1−ρ)Lα(j|i)A(i→j)=(1-ρ)L_α(j\,|\,i), where A denotes the transition kernel as in Rabiner 1989. Here LαL_α shrinks the add-one-smoothed transition estimate L toward the uniform null using the positive-part form of James et al. 1961: Lα(j|i)=1−αK−1+αL(j|i),α=(1−dfχ2)+,L_α(j\,|\,i)= 1-αK-1+α L(j\,|\,i), α= (1- dfχ^2 )^+, (6) where (z)+=max(z,0)(z)^+= (z,0), χ2χ^2 is Pearson’s statistic for the development destination counts against the uniform null, and dfdf is its degrees of freedom. Since [χ2]=dfE[χ^2]=df under the null, α=0α=0 whenever the observed statistic does not exceed its null expectation, causing the filter to reduce to Equation 5 rather than fitting a spurious transition table. We use this positive-part form as a closed-form shrinkage rule toward the uniform null; Appendix J reports its bootstrap stability. The estimate is closed-form and adds no tuned scalar, no training, and no deployment cost. 3.6 Prompt-Free Decontamination Operator When the perception prompt cannot be firewalled, as in legacy or third-party serving stacks, the same intervention grid yields a repair rather than a diagnosis. Writing q(∅)q^( ) for the clean posterior and q(b)q^(b) for the one contaminated by previous label b, we estimate a per-label log-offset and invert it, δb(y)=[logq(∅)(y)−logq(b)(y)],q^∝q(b)exp(δb). _b(y)=E\! [ q^( )(y)- q^(b)(y) ], q q^(b) ( _b). (7) The estimate is closed-form, needs no prompt access, no gradient training, and no extra SpeechLM calls. It requires only the label that the deployed system itself wrote into the prompt. We fit δb _b on one stratified half of the grid and apply it to the held-out half (Appendix I). Budget-matched endpoint calibration. A frozen SpeechLM’s label posterior carries its own class prior, which need not match the benchmark’s. We therefore optionally calibrate the terminal observation by the development prior ratio w(y)=π^gold(y)/π^pred(y)w(y)= π_gold(y)/ π_pred(y), the standard prior-adjustment correction for label shift (Saerens et al. 2002; Lipton et al. 2018); dividing a discriminative posterior by its class prior to obtain an HMM-compatible likelihood is the scaled-likelihood construction of hybrid speech recognition (Bourlard & Morgan 2012). This adds no search budget: five values of ρ are paired with two endpoint modes, forming ten pre-specified filter configurations. These are separate from the ten grounding-prompt candidates in Section 3.3. Both are selected using development data only and locked before evaluation, while α and w are estimated from development labels only. 4 Experiments In this section we carry out experiments to address the following questions: • Q1: Does a previous belief change what a frozen SpeechLM hears from fixed audio? See §4.2. • Q2: Where does history matter most? See §4.3. • Q3: Does decoupling improve streaming trajectories? See §4.4. • Q4: Which components drive the gain? See §4.5. • Q5: Is the result robust, and can the measurements themselves repair the failure? See §4.6. The appendix adds (1) per-metric tables, (2) mechanism diagnostics, (3) difficulty strata, (4) grid-size and (5) bootstrap studies, (6) prompt and baseline rules, and (7) additional stress tests. 4.1 Experimental Setup Benchmarks and models. We evaluate on two complementary benchmarks (Table 4, Appendix C). CREMA-D-Stream contains four-chunk same-speaker trajectories that either remain stable or switch emotion after two chunks, with disjoint development and evaluation speakers. HumDial-En converts the English emotion-trajectory task of HumDial-EIBench (Wang et al. 2026a) into causal chunks of human-recorded multi-turn speech, with disjoint source groups. We freeze Qwen2-Audio-7B-Instruct (Chu et al. 2024), Qwen2.5-Omni-7B-AWQ (Xu et al. 2025), Phi-4-multimodal-instruct (Abouelenin et al. 2025), and MiniCPM-o-2.6 (Yao et al. 2024), using temperature-zero decoding. Baselines and compute. We compare with direct history conditioning (DHC), independent chunk inference (ICI), confidence- and margin-gated revision (CGR/MGR), causal HMM filtering (C-HMM), and instantaneous/accumulated JSD revision (IJSR/AJSR). Each tunable baseline policy receives one scalar searched over the same ten-value development grid (Appendix D). All controlled policies replay the same posterior cache with one SpeechLM call per chunk, zero policy calls, and zero trainable parameters. Grounding-prompt search evaluates the fixed ten-candidate instruction pool once and is reported separately (Appendix G). Metrics and selection. Our primary metric is S-BAcc, averaging accuracy over true-change and stable steps; we additionally report step macro-F1, step accuracy, and final accuracy. Transition F1, inertia, volatility, and revision delay serve as mechanism diagnostics (Appendix K). All belief policies and ρ values are selected on development data using the same pre-specified ranking (S-BAcc, step macro-F1, Transition F1, step accuracy, final accuracy), while grounding prompts follow Section 3.3. All choices are locked before evaluation. We report full-set point estimates with B=1000B=1000 episode-bootstrap standard errors (Appendix P). 4.2 A1: Previous Beliefs Contaminate Current Perception Figure 3: Counterfactual previous-label intervention (CREMA-D-Stream, Qwen2-Audio). Eligible predictions captured by the wrong injected previous label (terracotta: injected label wins). A twentyfold spread shows that no uniform stickiness repairs the effect. Destinations: Figure 7. We sample 120 balanced CREMA-D-Stream development chunks (20 per emotion) and inject all six candidate previous labels, giving 720 injection conditions (600 pull-eligible); the asymmetry estimates below are chunk-level, not episode-level. As Figure 3 shows, changing only the injected previous label strongly shifts the output: accuracy falls from 72.50% without previous-label exposure to 30.42% under exposure, and 65.69% of predictions flip. Among wrong injected labels, 71.00% pull the output to the injected label, while audible evidence overrides one in only 23.68% of conditions where the control was correct; mean posterior JSD is 0.4823. The cost is visible at deployment scale on every backbone (Table 2; Figure 6, Appendix H): direct history conditioning loses up to 49 step-accuracy points, and EmoUpdate is highest in all eight settings. Instruction-only mitigation remains insufficient in this diagnostic: across all ten grounding candidates, seven of which explicitly forbid using the previous label, wrong-prior pull stays above 59.7% and rises with prior-blind accuracy (Pearson r=0.81r=0.81). This motivates removing historical belief from the perception call rather than merely instructing the model to ignore it (Figure 8, Appendix H). 4.3 A2: History Matters Exactly Where Perception Is Weak Table 1 localizes both the value and the danger of history across the three backbones on which the balanced injection was run. Call a chunk ambiguous when its prior-blind posterior margin falls below the development median. There, a correct previous label is worth +17.2+17.2 to +40.9+40.9 accuracy points on every backbone and benchmark, while a wrong one is adopted in 26.426.4–86.9%86.9\% of eligible cases. On easy chunks the policies become numerically indistinguishable because the audio already settles the label. Benefit and hazard live on the same subset, which is why history must enter through a filter rather than through the prompt. Replaying every policy on one shared cache confirms that EmoUpdate collects this benefit wherever the acoustic posterior supports it: +24.07+24.07 for Qwen2-Audio (p<10−4p<10^-4) and +8.66+8.66 for Phi-4M (p=0.019p=0.019) on HumDial-En, and +3.07+3.07 on CREMA-D-Stream (p=0.040p=0.040), against at most +1.28+1.28 on easy chunks in any setting. The gain is therefore not a uniform prompt effect. When the prior-blind posterior carries little usable acoustic information, as on CREMA-D-Stream with Phi-4M and MiniCPM-o, both below 25% chunkwise accuracy, the policy differences become correspondingly small; Appendix F reports the full stratification. Table 1: History is worth most exactly where perception is weakest. A chunk is ambiguous when its prior-blind posterior margin falls below the development median and easy when margin and duration are both above it, so the two groups are not complementary; thresholds are never tuned on test. Δ is the gain from supplying the correct previous label, pull the rate at which a wrong one is adopted, and Δ the gain over independent chunk inference on one shared posterior cache (exact paired McNemar p). All five difficulty groups are in Appendix F. Ambiguous chunks Easy chunks Data Model Δ Pull Δ p Δ Δ HumDial-En Qwen2-Audio +39.29 63.7 +24.07 <<1e–4 +0.00 +0.00 Phi-4M +27.27 30.3 +8.66 0.019 +5.88 +0.00 MiniCPM-o 2.6 +17.24 26.4 +2.74 0.125 +0.00 +0.00 CREMA-D-Stream Qwen2-Audio +36.07 86.9 +3.07 0.040 +4.35 +0.00 Phi-4M +25.86 49.0 +0.58 0.664 +42.86 +0.00 MiniCPM-o 2.6 +40.91 63.6 +0.63 0.749 +23.53 +1.28 4.4 A3: Decoupling Improves Streaming Trajectories Table 2 reports the complete controlled comparison; its Δ row is the strict per-metric envelope, so the reference values need not come from a single baseline. Table 2: Controlled comparison across four frozen SpeechLMs (%, point estimates with bootstrap standard errors; shared posterior cache, one call per chunk). Primary metrics first: state-balanced accuracy and class-balanced step macro-F1, then step accuracy and conventional final-turn accuracy. Bold: per-column best; shaded row: EmoUpdate; Δ : EmoUpdate minus the per-metric best controlled baseline. HumDial-En CREMA-D-Stream Model Method S-BAcc ↑ Step-MaF1 ↑ Step ↑ Final ↑ S-BAcc ↑ Step-MaF1 ↑ Step ↑ Final ↑ Qwen2-Audio DHC 1.43±0.63± 0.63 20.68±1.81± 1.81 20.00±1.83± 1.83 4.76±2.08± 2.08 18.28±2.18± 2.18 10.97±1.39± 1.39 20.70±2.17± 2.17 18.85±2.50± 2.50 ICI 29.14±1.94± 1.94 68.28±2.00± 2.00 60.95±2.15± 2.15 89.52±3.00± 3.00 70.33±2.61± 2.61 68.41±2.01± 2.01 69.57±2.20± 2.20 70.08±2.95± 2.95 CGR 21.71±1.36± 1.36 55.73±2.13± 2.13 53.97±2.27± 2.27 72.38±4.19± 4.19 70.00±2.63± 2.63 68.25±2.02± 2.02 69.57±2.21± 2.21 70.08±2.95± 2.95 MGR 24.57±1.25± 1.25 60.73±2.07± 2.07 57.14±2.17± 2.17 81.90±3.67± 3.67 70.25±2.63± 2.63 68.48±2.01± 2.01 69.88±2.21± 2.21 70.49±2.98± 2.98 C-HMM 29.43±2.54± 2.54 67.26±2.08± 2.08 60.00±2.34± 2.34 85.71±3.37± 3.37 70.33±2.61± 2.61 68.66±2.00± 2.00 69.98±2.20± 2.20 70.90±2.97± 2.97 IJSR 30.00±2.68± 2.68 67.81±2.05± 2.05 60.63±2.25± 2.25 87.62±3.16± 3.16 69.18±2.77± 2.77 67.15±2.27± 2.27 68.55±2.42± 2.42 67.21±3.10± 3.10 AJSR 23.43±1.31± 1.31 59.82±2.15± 2.15 55.87±2.26± 2.26 78.10±3.90± 3.90 70.66±2.73± 2.73 68.71±2.19± 2.19 69.98±2.34± 2.34 69.67±3.05± 3.05 EmoUpdate 70.57±4.16± 4.16 82.31±2.21± 2.21 82.54±2.02± 2.02 87.62±3.27± 3.27 74.02±2.58± 2.58 74.58±1.77± 1.77 76.23±2.06± 2.06 77.46±2.72± 2.72 Δ vs. best ctl. +40.57 +14.03 +21.59 −-1.90 +3.36 +5.87 +6.25 +6.56 Qwen2.5-Omni DHC 73.14±4.28± 4.28 76.63±2.44± 2.44 76.19±2.42± 2.42 80.95±3.81± 3.81 26.15±2.25± 2.25 26.77±1.43± 1.43 30.53±1.83± 1.83 24.18±2.70± 2.70 ICI 81.43±4.17± 4.17 86.14±2.03± 2.03 86.03±2.00± 2.00 89.52±2.93± 2.93 42.30±2.64± 2.64 42.21±1.80± 1.80 43.85±2.18± 2.18 48.77±3.11± 3.11 CGR 81.43±4.17± 4.17 85.90±2.06± 2.06 86.03±2.00± 2.00 89.52±2.93± 2.93 42.30±2.64± 2.64 42.21±1.80± 1.80 43.85±2.18± 2.18 48.77±3.11± 3.11 MGR 81.43±4.17± 4.17 86.14±2.03± 2.03 86.03±2.00± 2.00 89.52±2.93± 2.93 42.30±2.64± 2.64 42.21±1.80± 1.80 43.85±2.18± 2.18 48.77±3.11± 3.11 C-HMM 79.43±4.28± 4.28 85.54±2.08± 2.08 85.08±2.12± 2.12 89.52±2.93± 2.93 43.03±2.64± 2.64 42.21±1.79± 1.79 43.95±2.18± 2.18 47.95±3.09± 3.09 IJSR 79.71±4.41± 4.41 85.74±2.08± 2.08 85.40±2.12± 2.12 88.57±2.99± 2.99 42.87±2.66± 2.66 43.17±1.91± 1.91 44.57±2.23± 2.23 49.18±3.14± 3.14 AJSR 85.43±3.86± 3.86 86.49±2.08± 2.08 86.67±2.06± 2.06 89.52±2.93± 2.93 42.62±2.66± 2.66 42.67±1.83± 1.83 44.26±2.20± 2.20 50.00±3.18± 3.18 EmoUpdate 96.29±0.95± 0.95 90.66±1.74± 1.74 91.75±1.51± 1.51 91.43±2.67± 2.67 53.85±2.51± 2.51 52.45±1.78± 1.78 54.00±2.08± 2.08 67.21±3.07± 3.07 Δ vs. best ctl. +10.86 +4.18 +5.08 +1.90 +10.82 +9.28 +9.43 +17.21 Phi-4M DHC 14.00±3.70± 3.70 38.38±2.19± 2.19 36.51±2.26± 2.26 13.33±3.46± 3.46 16.64±2.21± 2.21 10.67±1.06± 1.06 14.24±1.57± 1.57 15.98±2.41± 2.41 ICI 21.14±1.46± 1.46 41.82±2.51± 2.51 43.81±2.30± 2.30 67.62±4.49± 4.49 15.98±2.12± 2.12 4.70±0.49± 0.49 16.39±2.01± 2.01 15.98±2.31± 2.31 CGR 18.57±1.41± 1.41 36.05±2.24± 2.24 40.95±2.17± 2.17 61.90±4.63± 4.63 15.98±2.12± 2.12 4.70±0.49± 0.49 16.39±2.01± 2.01 15.98±2.31± 2.31 MGR 21.14±1.46± 1.46 41.82±2.51± 2.51 43.81±2.30± 2.30 67.62±4.49± 4.49 15.98±2.12± 2.12 4.70±0.49± 0.49 16.39±2.01± 2.01 15.98±2.31± 2.31 C-HMM 20.86±1.45± 1.45 41.03±2.37± 2.37 43.49±2.27± 2.27 66.67±4.44± 4.44 15.98±2.12± 2.12 4.70±0.49± 0.49 16.39±2.01± 2.01 15.98±2.31± 2.31 IJSR 20.86±1.49± 1.49 41.65±2.53± 2.53 43.49±2.37± 2.37 66.67±4.56± 4.56 15.98±2.12± 2.12 4.70±0.49± 0.49 16.39±2.01± 2.01 15.98±2.31± 2.31 AJSR 19.71±1.42± 1.42 38.95±2.21± 2.21 42.22±2.27± 2.27 65.71±4.48± 4.48 15.98±2.12± 2.12 4.70±0.49± 0.49 16.39±2.01± 2.01 15.98±2.31± 2.31 EmoUpdate 90.86±2.22± 2.22 78.54±2.64± 2.64 82.22±2.25± 2.25 80.00±3.88± 3.88 18.93±2.18± 2.18 10.85±1.04± 1.04 19.88±1.99± 1.99 20.90±2.57± 2.57 Δ vs. best ctl. +69.71 +36.72 +38.41 +12.38 +2.30 +0.18 +3.48 +4.92 MiniCPM-o 2.6 DHC 30.00±5.54± 5.54 42.42±2.34± 2.34 42.86±2.22± 2.22 33.33±4.62± 4.62 18.69±2.19± 2.19 14.42±1.30± 1.30 20.39±2.06± 2.06 20.08±2.52± 2.52 ICI 79.71±4.12± 4.12 84.67±1.88± 1.88 84.13±1.99± 1.99 93.33±2.38± 2.38 20.08±2.20± 2.20 13.75±1.07± 1.07 22.23±2.08± 2.08 22.13±2.68± 2.68 CGR 79.71±4.12± 4.12 84.67±1.88± 1.88 84.13±1.99± 1.99 93.33±2.38± 2.38 20.08±2.20± 2.20 13.74±1.07± 1.07 22.23±2.08± 2.08 22.13±2.68± 2.68 MGR 79.43±4.21± 4.21 84.16±1.91± 1.91 83.81±2.03± 2.03 92.38±2.53± 2.53 19.92±2.20± 2.20 13.51±1.07± 1.07 22.03±2.08± 2.08 21.72±2.65± 2.65 C-HMM 77.71±4.32± 4.32 83.72±1.98± 1.98 83.17±2.11± 2.11 92.38±2.53± 2.53 20.16±2.19± 2.19 13.94±1.08± 1.08 22.34±2.08± 2.08 22.13±2.68± 2.68 IJSR 81.71±4.08± 4.08 84.57±1.90± 1.90 83.81±2.09± 2.09 92.38±2.53± 2.53 20.16±2.19± 2.19 13.82±1.07± 1.07 22.34±2.10± 2.10 22.13±2.68± 2.68 AJSR 79.14±4.24± 4.24 83.87±1.98± 1.98 83.49±2.11± 2.11 92.38±2.53± 2.53 19.92±2.22± 2.22 13.50±1.16± 1.16 22.03±2.12± 2.12 20.90±2.57± 2.57 EmoUpdate 91.43±2.83± 2.83 82.95±2.27± 2.27 85.08±1.97± 1.97 91.43±2.64± 2.64 24.18±1.97± 1.97 22.61±1.36± 1.36 25.20±1.65± 1.65 28.69±2.85± 2.85 Δ vs. best ctl. +9.71 −-1.72 +0.95 −-1.90 +4.02 +8.18 +2.87 +6.56 On HumDial-En, EmoUpdate leads on S-BAcc and step accuracy for every model. With Qwen2-Audio it gains +40.57+40.57 S-BAcc, +21.59+21.59 step accuracy, and +14.03+14.03 step macro-F1 over the controlled envelope. With Phi-4M it prevents a collapse, raising S-BAcc by 69.71 points. On CREMA-D-Stream, EmoUpdate wins all four reported metrics on every backbone, with its largest endpoint gain here (+17.21+17.21 final accuracy on Qwen2.5-Omni). Across both benchmarks, EmoUpdate is best on step accuracy and S-BAcc in 8/88/8 settings, on step macro-F1 in 7/87/8, and on final accuracy in 6/86/8. Appendix L additionally compares EmoUpdate with the published contextual model DVL-CER (Zha et al. 2025); EmoUpdate is better in all 24 trajectory cells. 4.5 A4: Grounding Enables Effective Causal Revision Table 3 separates grounding from filtering, applying the same deployed filter, including its asymmetric kernel and shrinkage, in both policy arms. The policy-only column therefore applies an asymmetry-aware filter to the contaminated observation. Replacing that observation with a prior-blind acoustic posterior is the larger effect in all eight settings. Causal filtering then performs history-aware revision on the grounded posterior and further improves S-BAcc in seven of eight settings. On HumDial-En with Qwen2-Audio, the same filter contributes +8.29+8.29 S-BAcc before grounding and +16.86+16.86 after it, showing that temporal revision benefits from a cleaner observation. Table 3: Component ablation on state-balanced accuracy (%, point estimates). All arms replay the same episodes. “Policy” applies the deployed evidence-shrunk asymmetric filter (Eq. 6) to the contaminated seed observation; “Ground” applies chunkwise updates to the prior-blind observation; and “Both” is the full firewalled inference path. The two stages of this path are complementary: grounding restores the current-audio observation, while causal filtering incorporates temporal state only after perception has been decoupled from historical belief. HumDial-En CREMA-D-Stream Model Neither Policy Ground Both Δ . Δ . Neither Policy Ground Both Δ . Δ . Qwen2-Audio 29.14 37.43 53.71 70.57 +24.57 +16.86 70.33 68.69 74.75 74.02 +4.43 −-0.74 Qwen2.5-Omni 81.43 86.29 91.43 96.29 +10.00 +4.86 42.30 45.00 50.66 53.85 +8.36 +3.20 Phi-4M 21.14 28.57 84.29 90.86 +63.14 +6.57 15.98 16.39 18.69 18.93 +2.70 +0.25 MiniCPM-o 2.6 79.71 80.57 90.29 91.43 +10.57 +1.14 20.08 21.31 24.02 24.18 +3.93 +0.16 Figure 4 shows where the gains arise: seed-posterior policies collapse at the second turn on Qwen2-Audio and Phi-4M, the first turn at which a previous belief exists, while EmoUpdate stays nearly flat. Where the backbone is already strong, the curves separate less, consistent with contamination rather than capacity being the limit. Figure 4: Causal step accuracy over the streaming trajectory, all backbones and controlled baselines. Shading is EmoUpdate’s margin over the best baseline (leads / trails). The CREMA-D-Stream row uses per-panel y-ranges; read the ticks first. 4.6 A5: Robustness and Measurement-Parameterized Repairs Figure 5: The locked stay-probability is not a delicate choice. Band: every ρ within 2 S-BAcc points of the panel best (dot). Dashed line: the ρ locked on development data, never read off this plot; inside the band in six of eight settings and within 2.9 points of the optimum in all eight. α is the coefficient of Equation 6. The four backbones react differently to history, yet the same design improves step accuracy in every setting under an unchanged resource contract. The 51-point ρ replay (Figure 5) shows a broad plateau that every budget prefix (Table 21) lands on. The shrinkage rule adapts to the observed transition structure: HumDial-En development evidence gives α=0.87α=0.87, lifting Qwen2-Audio’s step accuracy from 69.8 to 82.2, whereas CREMA-D-Stream, whose shifts are uniform by construction, returns α=0.00α=0.00. In this case the transition prior reduces to the symmetric form of Equation 5, avoiding an unsupported asymmetric transition structure. Cross-validated on the balanced intervention grid, the operator of Section 3.6 lifts exposed accuracy from 30.42 to 53.89 (control 72.50) and reduces wrong-prior pull from 71.00 to 12.17 (Appendix I), without prompt access, training, or additional SpeechLM calls. 5 Conclusion We identified previous-belief contamination: a streaming SpeechLM’s own previous prediction can alter its perception of identical current audio, with strong label asymmetry and the greatest risk where acoustic perception is weakest. EmoUpdate addresses this failure at its source by keeping current perception prior-blind and introducing history only afterward through an evidence-shrunk causal filter, with a training-free decontamination fallback when firewalling is unavailable. Historical state should update a grounded observation, never help create it. Ethics Statement Emotion recognition is culturally and contextually uncertain, and systems built on it should not be used for clinical diagnosis, hiring decisions, or consequential surveillance. PBC adds a specific hazard: a streaming system can amplify an erroneous earlier label into a confident trajectory. Deployments should expose uncertainty and permit correction rather than treat filtered beliefs as facts. All source corpora retain their original licenses, and constructed trajectories introduce no generated emotion labels (Appendix P). Reproducibility Statement Every claim maps to a released artifact: episode manifests with data hashes, perception caches, selection locks that freeze candidate order and chosen parameters before evaluation, per-step predictions, bootstrap summaries, plot-source tables, and a deterministic figure script. Appendix P describes the six-stage locked reproduction order; Appendices C and D give dataset construction, leakage controls, and exact baseline decision rules. References Abouelenin et al. (2025) Abdelrahman Abouelenin, Atabak Ashfaq, Adam Atkinson, Hany Awadalla, Nguyen Bach, Jianmin Bao, Alon Benhaim, Martin Cai, Vishrav Chaudhary, Congcong Chen, et al. Phi-4-mini technical report: Compact yet powerful multimodal language models via mixture-of-loras. arXiv preprint arXiv:2503.01743, 2025. Arora et al. (2022) Kushal Arora, Layla El Asri, Hareesh Bahuleyan, and Jackie Chi Kit Cheung. Why exposure bias matters: An imitation learning perspective of error accumulation in language generation. In Findings of the Association for Computational Linguistics: ACL 2022, p. 700–710, 2022. Baevski et al. (2020) Alexei Baevski, Yuhao Zhou, Abdelrahman Mohamed, and Michael Auli. wav2vec 2.0: A framework for self-supervised learning of speech representations. Advances in neural information processing systems, 33:12449–12460, 2020. Bengio et al. (2015) Samy Bengio, Oriol Vinyals, Navdeep Jaitly, and Noam Shazeer. Scheduled sampling for sequence prediction with recurrent neural networks. Advances in neural information processing systems, 28, 2015. Bourlard & Morgan (2012) Herve A Bourlard and Nelson Morgan. Connectionist speech recognition: a hybrid approach. Springer Science & Business Media, 2012. Busso et al. (2008) Carlos Busso, Murtaza Bulut, Chi-Chun Lee, Abe Kazemzadeh, Emily Mower, Samuel Kim, Jeannette N Chang, Sungbok Lee, and Shrikanth S Narayanan. Iemocap: Interactive emotional dyadic motion capture database. Language resources and evaluation, 42(4):335–359, 2008. Cao et al. (2014) Houwei Cao, David G Cooper, Michael K Keutmann, Ruben C Gur, Ani Nenkova, and Ragini Verma. Crema-d: Crowd-sourced emotional multimodal actors dataset. IEEE transactions on affective computing, 5(4):377–390, 2014. Chen et al. (2022) Sanyuan Chen, Chengyi Wang, Zhengyang Chen, Yu Wu, Shujie Liu, Zhuo Chen, Jinyu Li, Naoyuki Kanda, Takuya Yoshioka, Xiong Xiao, et al. Wavlm: Large-scale self-supervised pre-training for full stack speech processing. IEEE Journal of Selected Topics in Signal Processing, 16(6):1505–1518, 2022. Chu et al. (2024) Yunfei Chu, Jin Xu, Qian Yang, Haojie Wei, Xipin Wei, Zhifang Guo, Yichong Leng, Yuanjun Lv, Jinzheng He, Junyang Lin, et al. Qwen2-audio technical report. arXiv preprint arXiv:2407.10759, 2024. Gong et al. (2024) Yuan Gong, Hongyin Luo, Alexander Liu, Leonid Karlinsky, and James R Glass. Listen, think, and understand. In International Conference on Learning Representations, volume 2024, p. 18516–18545, 2024. Goyal et al. (2024) Sachin Goyal, Christina Baek, J Zico Kolter, and Aditi Raghunathan. Context-parametric inversion: Why instruction finetuning can worsen context reliance. arXiv preprint arXiv:2410.10796, 2024. Guo et al. (2017) Chuan Guo, Geoff Pleiss, Yu Sun, and Kilian Q Weinberger. On calibration of modern neural networks. In International conference on machine learning, p. 1321–1330. PMLR, 2017. Guo et al. (2026) Pengze Guo, Jingxi Liang, Zhiwen Xie, Qifeng Wang, and Derek F Wong. Emos: A high-fidelity multimodal benchmark for fine-grained streaming emotional understanding. In Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), p. 39074–39089, 2026. Hsu et al. (2021) Wei-Ning Hsu, Benjamin Bolte, Yao-Hung Hubert Tsai, Kushal Lakhotia, Ruslan Salakhutdinov, and Abdelrahman Mohamed. Hubert: Self-supervised speech representation learning by masked prediction of hidden units. IEEE/ACM transactions on audio, speech, and language processing, 29:3451–3460, 2021. Huang et al. (2024) Jie Huang, Xinyun Chen, Swaroop Mishra, Huaixiu Steven Zheng, Adams Yu, Xinying Song, and Denny Zhou. Large language models cannot self-correct reasoning yet. In International conference on learning representations, volume 2024, p. 32808–32824, 2024. James et al. (1961) William James, Charles Stein, et al. Estimation with quadratic loss. In Proceedings of the fourth Berkeley symposium on mathematical statistics and probability, volume 1, p. 361–379. University of California Press, 1961. Kaplan & Wermter (2026) Burak Can Kaplan and Stefan Wermter. Scope: Shift-aware speaker-conditioned priors for emotion recognition in conversations. arXiv preprint arXiv:2607.20445, 2026. Lamb et al. (2016) Alex M Lamb, Anirudh Goyal Alias Parth Goyal, Ying Zhang, Saizheng Zhang, Aaron C Courville, and Yoshua Bengio. Professor forcing: A new algorithm for training recurrent networks. Advances in neural information processing systems, 29, 2016. Lipton et al. (2018) Zachary Lipton, Yu-Xiang Wang, and Alexander Smola. Detecting and correcting for label shift with black box predictors. In International conference on machine learning, p. 3122–3130. PMLR, 2018. Liu et al. (2024a) Nelson F Liu, Kevin Lin, John Hewitt, Ashwin Paranjape, Michele Bevilacqua, Fabio Petroni, and Percy Liang. Lost in the middle: How language models use long contexts. Transactions of the association for computational linguistics, 12:157–173, 2024a. Liu et al. (2024b) Yingjian Liu, Jiang Li, Xiaoping Wang, and Zhigang Zeng. Emotionic: emotional inertia and contagion-driven dependency modeling for emotion recognition in conversation. Science China Information Sciences, 67(8):182103, 2024b. Lotfian & Busso (2017) Reza Lotfian and Carlos Busso. Building naturalistic emotionally balanced speech corpus by retrieving emotional speech from existing podcast recordings. IEEE Transactions on Affective Computing, 10(4):471–483, 2017. Lu et al. (2022) Yao Lu, Max Bartolo, Alastair Moore, Sebastian Riedel, and Pontus Stenetorp. Fantastically ordered prompts and where to find them: Overcoming few-shot prompt order sensitivity. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), p. 8086–8098, 2022. Meng et al. (2022) Kevin Meng, David Bau, Alex Andonian, and Yonatan Belinkov. Locating and editing factual associations in gpt. Advances in neural information processing systems, 35:17359–17372, 2022. Pang et al. (2026) Jiacheng Pang, Ashutosh Chaubey, and Mohammad Soleymani. Do audio llms listen or read? analyzing and mitigating paralinguistic failures with voxparadox. arXiv preprint arXiv:2605.27772, 2026. Perez et al. (2022) Ethan Perez, Sam Ringer, Kamilė Lukošiūtė, Karina Nguyen, Edwin Chen, Scott Heiner, Craig Pettit, C Olsson, S Kundu, S Kadavath, et al. Discovering language model behaviors with model-written evaluations. arxiv. arXiv preprint arXiv:2212.09251, 2022. Poria et al. (2019a) Soujanya Poria, Devamanyu Hazarika, Navonil Majumder, Gautam Naik, Erik Cambria, and Rada Mihalcea. Meld: A multimodal multi-party dataset for emotion recognition in conversations. In Proceedings of the 57th annual meeting of the association for computational linguistics, p. 527–536, 2019a. Poria et al. (2019b) Soujanya Poria, Navonil Majumder, Rada Mihalcea, and Eduard Hovy. Emotion recognition in conversation: Research challenges, datasets, and recent advances. IEEE access, 7:100943–100953, 2019b. Rabiner (1989) Lawrence R Rabiner. A tutorial on hidden markov models and selected applications in speech recognition. Proceedings of the IEEE, 77(2):257–286, 1989. Ranzato et al. (2015) Marc’Aurelio Ranzato, Sumit Chopra, Michael Auli, and Wojciech Zaremba. Sequence level training with recurrent neural networks. arXiv preprint arXiv:1511.06732, 2015. Saerens et al. (2002) Marco Saerens, Patrice Latinne, and Christine Decaestecker. Adjusting the outputs of a classifier to new a priori probabilities: a simple procedure. Neural computation, 14(1):21–41, 2002. Schmidt (2019) Florian Schmidt. Generalization in generation: A closer look at exposure bias. In Proceedings of the 3rd Workshop on Neural Generation and Translation, p. 157–167, 2019. Sharma et al. (2024) Mrinank Sharma, Meg Tong, Tomek Korbak, David Duvenaud, Amanda Askell, Sam Bowman, Esin Durmus, Zac Hatfield-Dodds, Scott Johnston, Shauna Kravec, et al. Towards understanding sycophancy in language models. In International Conference on Learning Representations, volume 2024, p. 110–144, 2024. Shen et al. (2024) Siyuan Shen, Yu Gao, Feng Liu, Hanyang Wang, and Aimin Zhou. Emotion neural transducer for fine-grained speech emotion recognition. In ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), p. 10111–10115. IEEE, 2024. Song et al. (2026) Zhiyuan Song, Weici Zhao, Yang Xiao, Suhao Yu, Cheng Zhu, and Jiatao Gu. Learning when to think while listening in large audio-language models. arXiv preprint arXiv:2605.27190, 2026. Tang et al. (2024) Changli Tang, Wenyi Yu, Guangzhi Sun, Xianzhao Chen, Tian Tan, Wei Li, Lu Lu, Zejun Ma, and Chao Zhang. Salmonn: Towards generic hearing abilities for large language models. In International Conference on Learning Representations, volume 2024, p. 16607–16629, 2024. Turpin et al. (2023) Miles Turpin, Julian Michael, Ethan Perez, and Samuel Bowman. Language models don’t always say what they think: Unfaithful explanations in chain-of-thought prompting. Advances in Neural Information Processing Systems, 36:74952–74965, 2023. Vig et al. (2020) Jesse Vig, Sebastian Gehrmann, Yonatan Belinkov, Sharon Qian, Daniel Nevo, Yaron Singer, and Stuart Shieber. Investigating gender bias in language models using causal mediation analysis. Advances in neural information processing systems, 33:12388–12401, 2020. Wang & Sennrich (2020) Chaojun Wang and Rico Sennrich. On exposure bias, hallucination and domain shift in neural machine translation. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, p. 3544–3552, 2020. Wang et al. (2026a) Shuiyuan Wang, Zhixian Zhao, Hongfei Xue, Chengyou Wang, Shuai Wang, Hui Bu, Xin Xu, and Lei Xie. Humdial-eibench: A human-recorded multi-turn emotional intelligence benchmark for audio language models. arXiv preprint arXiv:2604.11594, 2026a. Wang et al. (2022) Xuezhi Wang, Jason Wei, Dale Schuurmans, Quoc Le, Ed Chi, Sharan Narang, Aakanksha Chowdhery, and Denny Zhou. Self-consistency improves chain of thought reasoning in language models. arXiv preprint arXiv:2203.11171, 2022. Wang et al. (2026b) Yancheng Wang, Osama Hanna, Ruiming Xie, Xianfeng Rui, Maohao Shen, Xuedong Zhang, Christian Fuegen, Jilong Wu, Debjyoti Paul, Arthur Guo, et al. Vowelprompt: Hearing speech emotions from text via vowel-level prosodic augmentation. arXiv preprint arXiv:2602.06270, 2026b. Xu et al. (2025) Jin Xu, Zhifang Guo, Jinzheng He, Hangrui Hu, Ting He, Shuai Bai, Keqin Chen, Jialin Wang, Yang Fan, Kai Dang, Bin Zhang, Xiong Wang, Yunfei Chu, and Junyang Lin. Qwen2.5-omni technical report, 2025. URL https://arxiv.org/abs/2503.20215. Xu et al. (2024) Rongwu Xu, Brian Lin, Shujian Yang, Tianqi Zhang, Weiyan Shi, Tianwei Zhang, Zhixuan Fang, Wei Xu, and Han Qiu. The earth is flat because…: Investigating llms’ belief towards misinformation via persuasive conversation. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), p. 16259–16303, 2024. Yao et al. (2024) Yuan Yao, Tianyu Yu, Ao Zhang, Chongyi Wang, Junbo Cui, Hongji Zhu, Tianchi Cai, Haoyu Li, Weilin Zhao, Zhihui He, et al. Minicpm-v: A gpt-4v level mllm on your phone. arXiv preprint arXiv:2408.01800, 2024. Zha et al. (2025) Xupeng Zha, Huan Zhao, Guanghui Ye, and Zixing Zhang. Dual-view learning for conversational emotion recognition through context and emotion-shift modeling. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 39, p. 25823–25831, 2025. Zhao et al. (2021) Zihao Zhao, Eric Wallace, Shi Feng, Dan Klein, and Sameer Singh. Calibrate before use: Improving few-shot performance of language models. In International conference on machine learning, p. 12697–12706. Pmlr, 2021. Contents 1 Introduction 2 Related Work 3 Method 3.1 Streaming Emotion Belief Revision 3.2 Counterfactual Previous-Belief Intervention 3.3 Acoustic History Firewall 3.4 Causal Belief Filter 3.5 Evidence-Shrunk Transition Prior 3.6 Prompt-Free Decontamination Operator 4 Experiments 4.1 Experimental Setup 4.2 A1: Previous Beliefs Contaminate Current Perception 4.3 A2: History Matters Exactly Where Perception Is Weak 4.4 A3: Decoupling Improves Streaming Trajectories 4.5 A4: Grounding Enables Effective Causal Revision 4.6 A5: Robustness and Measurement-Parameterized Repairs 5 Conclusion References A Extended Related Work B Metric Definitions C Dataset Construction and Leakage Controls D Controlled Baseline Definitions E Selected Human-Readable Prompts F Difficulty-Stratified Analysis G Compute and Fairness Audit H Additional Intervention Results I Measurement-Based Decontamination J Asymmetry-Aware Transition Prior K Conditional Volatility L Published-Method Adaptation M Cross-Benchmark Component Analysis N Selection-Budget and Parameter Sensitivity O Additional Diagnostic Analyses P Reproducibility and Responsible Use Appendix A Extended Related Work Exposure bias and error accumulation. Sequence models trained with teacher forcing but run on their own predictions accumulate error, the failure known as exposure bias. Scheduled sampling anneals ground-truth tokens toward model samples (Bengio et al. 2015), sequence-level training optimizes the evaluation metric directly (Ranzato et al. 2015), and professor forcing matches the free-running and teacher-forced dynamics adversarially (Lamb et al. 2016). Later analyses tie the effect to hallucination and domain shift in translation (Wang & Sennrich 2020), question how large it is for open-ended generation (Schmidt 2019), and recast it as compounding error in imitation learning (Arora et al. 2022). Every one of these remedies acts on how the next token is produced while the input representation is held fixed. PBC is the complementary failure: the input representation itself is rewritten by the previous output, so the same audio yields a different posterior. This is why our repair is a firewall on the perception call rather than a decoding rule, and why the diagnosis needs a counterfactual rather than a held-out split. Language models conditioned on asserted beliefs. Prompted LLMs are movable by what the context asserts. Sycophancy makes assistants agree with a stated user position (Sharma et al. 2024), an effect that scales with model-written evaluation pressure (Perez et al. 2022), and persuasive conversation can overturn a model’s factual belief (Xu et al. 2024). Stated reasoning need not be the operative cause (Turpin et al. 2023), and models do not reliably repair their own errors when asked (Huang et al. 2024); aggregating independent samples helps precisely because a single conditioned trajectory is unreliable (Wang et al. 2022). Purely positional factors move predictions as well: majority-label and recency bias in few-shot prompts (Zhao et al. 2021), example ordering (Lu et al. 2022), and the position of relevant evidence in a long context (Liu et al. 2024a). Our contaminating variable is narrower and recurrent: a single label the system itself emitted one chunk earlier, re-entering the prompt at every step. Causal intervention as an analysis tool. Holding an input fixed and editing one internal or textual variable is the standard way to move from correlation to mechanism in NLP analysis: causal mediation for attribute bias (Vig et al. 2020) and activation-level localization of factual associations (Meng et al. 2022). Our protocol applies the same logic at the prompt surface of a frozen SpeechLM: audio, instruction, and decoding are fixed and only the injected previous label varies. Speech representations, corpora, and audio-language models. Self-supervised encoders provide the acoustic front end most emotion systems build on (Baevski et al. 2020; Hsu et al. 2021; Chen et al. 2022). Emotion corpora span acted dyads (Busso et al. 2008), naturalistic podcasts (Lotfian & Busso 2017), and multi-party dialogue (Poria et al. 2019a), with conversational emotion recognition surveyed in Poria et al. 2019b. Instruction-following audio-language models extend this to open-ended listening (Tang et al. 2024; Gong et al. 2024; Chu et al. 2024; Xu et al. 2025). We hold such a model frozen and treat its posterior as the observation of a filter rather than as a final answer; the endpoint prior correction of Section 3.6 is a label-shift adjustment, distinct from confidence calibration in the temperature-scaling sense (Guo et al. 2017). Appendix B Metric Definitions Let ct=[yt≠yt−1]c_t=1[y_t≠ y_t-1] and c^t=[y^t≠y^t−1] c_t=1[ y_t≠ y_t-1]. A transition true positive requires ct=1c_t=1, c^t=1 c_t=1, and y^t=yt y_t=y_t. A predicted change in a stable region is a false positive; a true shift without a correct landing is a false negative. TF1 is the resulting F1 score. Change-state accuracy is Pr(y^t=yt∣ct=1) ( y_t=y_t c_t=1), stable-state accuracy is Pr(y^t=yt∣ct=0) ( y_t=y_t c_t=0), and S-BAcc is their arithmetic mean. Inertia is measured only when the previous prediction was correct: among steps with ct=1c_t=1 and y^t−1=yt−1 y_t-1=y_t-1, it is the fraction for which the policy does not change. Volatility is also conditioned on a previously correct prediction: among stable steps with y^t−1=yt−1=yt y_t-1=y_t-1=y_t, it is the fraction for which y^t≠yt y_t≠ y_t. This conditioning prevents already-wrong states from being counted as evidence of either desirable stability or destructive revision. Appendix C Dataset Construction and Leakage Controls CREMA-D-Stream. CREMA-D-Stream is constructed for this work from the original CREMA-D corpus (Cao et al. 2014), a collection of acted utterances recorded from professional actors over a fixed sentence set. We reorganize its utterances into four-chunk same-speaker streaming trajectories for causal trajectory evaluation. Each trajectory is assembled from four unique source utterances and contains a single speaker. A stable trajectory has the form (A,A,A,A)(A,A,A,A), while a change trajectory has the form (A,A,B,B)(A,A,B,B) with A≠BA≠ B. Each development speaker contributes two trajectories of each type; the evaluation split is constructed identically from disjoint speakers. The resulting development and evaluation sets contain 120 and 244 episodes, respectively. Across both splits, 1,456 source clips are selected without reuse. The 30 development and 61 evaluation speakers have zero overlap. HumDial-En. We preserve each Task-1 dialogue as an episode and expose the human-recorded non-query turns in their original temporal order. No future turn, reference answer, emotion cause, or Task-4 conflict annotation is supplied to the model. Development uses source group 0007; evaluation uses groups 0006, 0008, and 0010. This yields 45 development and 105 evaluation episodes with no source-group overlap. Task 4 remains separate from the main trajectory evaluation and is used only as an acoustic–semantic diagnostic. Table 4: Construction and leakage audit; each audio file appears in at most one trajectory. Property CREMA-D-Stream HumDial-En Episode source Same-speaker clips Original dialogue order Temporal pattern Stable or one shift Naturally occurring shifts Split key Speaker identity Released source group Dev/eval episodes 120 / 244 45 / 105 Dev/eval chunks 480 / 976 135 / 315 Label ontology 6 7 Dev/eval gold shifts 60 / 122 75 / 175 Rev./Keep (dev) 180 / 300 120 / 15 Rev./Keep (eval.) 366 / 610 280 / 35 Cross-split key overlap 0 0 Source-audio reuse None None across source groups Future information Never exposed Never exposed Appendix D Controlled Baseline Definitions Table 5 gives the exact causal decision rules used for the controlled envelope. Except for DHC, which deliberately tests direct history exposure, each rule operates on a cached current-chunk posterior. A threshold is chosen using development episodes only and is then locked before evaluation. Table 5: Controlled baseline rules (τ: development-selected scalar). Acronym Name Causal update rule DHC Direct history conditioning Put y^t−1 y_t-1 in the SpeechLM prompt ICI Independent chunk inference y^t=argmaxyqt(y) y_t= _yq_t(y) CGR Confidence-gated revision Revise if maxyqt(y)≥τ _yq_t(y)≥τ MGR Margin-gated revision Revise if qt(y∗)−qt(y^t−1)≥τq_t(y^*)-q_t( y_t-1)≥τ C-HMM Causal HMM Symmetric Markov prediction followed by Bayes update IJSR Instantaneous JSD revision Revise if DJS(qt,qt−1)≥τD_ JS(q_t,q_t-1)≥τ AJSR Accumulated JSD revision Accumulate confidence-weighted JSD to τ All seven systems are causal and use at most one frozen-SpeechLM call per chunk. CGR, MGR, C-HMM, IJSR, and AJSR receive the same ten-candidate scalar budget. DHC and ICI have no tunable scalar. These controls are deliberately simple because they isolate the two stages of the firewalled inference path; the separate DVL-CER adaptation in Appendix L provides a comparison to a recent published contextual model under its necessarily different training contract. Appendix E Selected Human-Readable Prompts The grounding prompt is chosen on development data only, under a rule fixed before evaluation: candidates whose final accuracy falls more than one percentage point below the seed prompt are rejected, and the remaining candidates are ranked lexicographically by S-BAcc, then Transition F1, step accuracy, and final accuracy, with grid order breaking exact ties. The pool contains ten candidates per setting. The same ten candidate families are fixed across the four backbones of each benchmark; development-time selection is performed separately for each model–benchmark setting, without redesigning the candidate pool. This rule is distinct from the belief-policy protocol of Section 4.1, which additionally uses step macro-F1 as its first tie-breaker; both are recorded in the released selection locks. Table 6 lists the selected candidate for each setting. The candidate pool always contains ten prompts. The exact prompt text, hashes, and development metrics are stored in the released selection locks. Table 6: Selected prior-blind prompt per setting. Benchmark Model Selected candidate CREMA-D-Stream Qwen2-Audio acoustic_first_no_copy CREMA-D-Stream Qwen2.5-Omni acoustic_first_original CREMA-D-Stream Phi-4M current_audio_contract CREMA-D-Stream MiniCPM-o 2.6 prior_deleted_check_short HumDial-En Qwen2-Audio seed_acoustic_short HumDial-En Qwen2.5-Omni acoustic_over_semantic_short HumDial-En Phi-4M prior_deleted_check_short HumDial-En MiniCPM-o 2.6 full_utterance_short For example, the selected Qwen2-Audio CREMA-D-Stream prompt is: Listen only to the CURRENT speech audio. Do not copy any previous label or label mentioned in history. Choose the emotion supported by audible pitch, energy, rhythm, pace, vocal tension, and voice quality. Reply with exactly one lowercase label: neutral, happy, sad, angry, fearful, or disgust. The history firewall is also enforced by the input wrapper: no previous label, belief distribution, prior utterance, filename, or speaker identifier is passed to the SpeechLM perception call. Five additional selected prompt bodies are listed below. Each candidate follows the same prior-blind perception contract, with the output ontology instantiated for the target benchmark. CREMA-D-Stream, Qwen2.5-Omni (acoustic_first_original). Listen only to the CURRENT speech audio. Identify emotion from audible pitch contour, loudness, tempo, rhythm, pauses, vocal tension, and voice quality. Do not use sentence meaning, history, or a previous label. Reply with exactly one lowercase label: neutral, happy, sad, angry, fearful, or disgust. CREMA-D-Stream, Phi-4M (current_audio_contract). The answer must depend only on the CURRENT speech audio. Previous beliefs, filenames, and speaker IDs are not evidence. Judge vocal emotion from pitch, loudness, tempo, pauses, rhythm, tension, and voice quality. Reply with exactly one lowercase label: neutral, happy, sad, angry, fearful, or disgust. HumDial-En, Qwen2-Audio (seed_acoustic_short). Listen only to the CURRENT attached speech audio with no history or previous label. Judge how the sentence is spoken from pitch, energy, rhythm, pace, and voice quality. Reply with exactly one lowercase HumDial-En label and nothing else. HumDial-En, Qwen2.5-Omni (acoustic_over_semantic_short). Listen to the CURRENT speech audio and classify how it is spoken. Prefer vocal prosody over literal sentence meaning or any previous belief. Use pitch, energy, timing, tension, and voice quality. Reply with exactly one lowercase HumDial-En label. HumDial-En, Phi-4M (prior_deleted_check_short). Listen only to the CURRENT speech audio. Choose the label you would choose if every previous-belief word were deleted from the prompt. Use pitch, loudness, tempo, rhythm, pauses, vocal tension, and voice quality. Reply with exactly one lowercase HumDial-En label. Table 7 lists the ten families the pool spans. Table 7: Ten prompt families in the fixed development pool. Candidate family Controlled design intent seed_acoustic_short Minimal current-audio instruction acoustic_first_original Enumerate prosodic evidence prior_distractor Mark prior labels as distractors acoustic_first_no_copy Explicitly forbid label copying current_audio_contract Restrict all evidence to current audio acoustic_over_semantic Prefer delivery under modality conflict label_anchor_guard State that ontology labels are not evidence full_utterance Require whole-chunk vocal judgment prior_deleted_check Counterfactually delete prior words compact_prototypes Supply short acoustic class prototypes Appendix F Difficulty-Stratified Analysis Table 8 expands Table 1 to all five difficulty groups. The strongest separation occurs on HumDial-En and on Qwen2-Audio CREMA-D-Stream, where the grounded observation supplies enough acoustic evidence for causal revision to act on. On the lower-information CREMA-D-Stream posteriors (19.6%19.6\% and 24.6%24.6\% chunkwise accuracy) margin-based difficulty is less discriminative and the policy differences are correspondingly small (p≥0.42p≥ 0.42); the same regime explains the large correct-prior gains on the nominally “easy” CREMA-D-Stream groups (+42.86+42.86 for Phi-4M, +23.53+23.53 for MiniCPM-o), since a high margin there marks confident errors rather than easy audio, while the policy gap stays at most 1.281.28 points. Exact group sizes and paired tests are reported for every group, including the two smaller short-and-ambiguous subsets on HumDial-En (N=45N=45 for Phi-4M, N=54N=54 for MiniCPM-o). Table 8: Complete difficulty-stratified analysis for the three backbones on which the balanced injection was run. Development-set medians define the groups: short is duration below the median, ambiguous is prior-blind posterior margin below the median, hard is both at once, and easy is neither. Left: prior intervention on the balanced sample. Right: all policies replayed on one shared posterior cache, with exact paired McNemar p against independent chunk inference. Prior intervention Shared-cache policy Data Model Chunks N No prior Correct Pull N ICI EmoUpdate Δ p HumDial-En Qwen2-Audio all 70 81.43 98.57 36.19 315 70.16 82.54 +12.38 <<1e–4 short 32 71.88 100.00 41.15 124 68.55 85.48 +16.94 <<1e–4 ambiguous 28 57.14 96.43 63.69 162 47.53 71.60 +24.07 <<1e–4 hard 14 42.86 100.00 67.86 69 49.28 79.71 +30.43 <<1e–4 easy 24 100.00 100.00 15.97 98 94.90 94.90 +0.00 1.00 Phi-4M all 70 77.14 92.86 18.33 315 78.73 82.22 +3.49 0.019 short 32 90.62 93.75 15.10 124 82.26 83.87 +1.61 0.688 ambiguous 33 57.58 84.85 30.30 127 58.27 66.93 +8.66 0.019 hard 12 83.33 83.33 29.17 45 64.44 68.89 +4.44 0.688 easy 17 94.12 100.00 8.82 109 92.66 92.66 +0.00 1.00 MiniCPM-o 2.6 all 70 92.86 100.00 14.76 315 83.81 85.08 +1.27 0.125 short 32 96.88 100.00 13.54 124 85.48 86.29 +0.81 1.00 ambiguous 29 82.76 100.00 26.44 146 69.18 71.92 +2.74 0.125 hard 11 90.91 100.00 25.76 54 70.37 72.22 +1.85 1.00 easy 20 100.00 100.00 5.83 99 95.96 95.96 +0.00 1.00 CREMA-D-Stream Qwen2-Audio all 120 80.00 100.00 85.50 976 74.69 76.23 +1.54 0.040 short 66 86.36 100.00 90.61 472 72.67 75.85 +3.18 0.006 ambiguous 61 63.93 100.00 86.89 488 60.45 63.52 +3.07 0.040 hard 30 73.33 100.00 93.33 230 55.22 61.74 +6.52 0.006 easy 23 95.65 100.00 77.39 246 88.62 88.62 +0.00 1.00 Phi-4M all 120 20.00 45.83 46.00 976 19.57 19.88 +0.31 0.664 short 66 22.73 42.42 41.52 472 20.76 20.13 -0.64 0.508 ambiguous 58 20.69 46.55 48.97 521 18.62 19.19 +0.58 0.664 hard 32 12.50 40.62 45.00 241 20.75 19.50 -1.24 0.508 easy 28 3.57 46.43 49.29 224 20.54 20.54 +0.00 1.00 MiniCPM-o 2.6 all 120 29.17 65.00 59.00 976 24.59 25.20 +0.61 0.581 short 66 31.82 65.15 56.97 472 23.31 22.46 -0.85 0.627 ambiguous 66 25.76 66.67 63.64 477 21.80 22.43 +0.63 0.749 hard 29 27.59 62.07 63.45 208 22.60 20.67 -1.92 0.424 easy 17 29.41 52.94 56.47 235 31.06 32.34 +1.28 0.648 Stable versus shift steps. Table 9 decomposes the same shared-cache replay by whether the true state changed at that step. To distinguish effective revision from simply revising more often, stable and true-shift steps are reported separately. On HumDial-En, shift accuracy improves on all three backbones, while stable accuracy improves on two and is preserved on the third. Stable accuracy rises by up to 14.2914.29 points, shift accuracy by up to 19.4319.43, and the false-switch rate falls by 2.862.86–11.4311.43 points. On CREMA-D-Stream, where transitions are uniform by construction and the shrinkage estimator returns α=0α=0, the filter behaves conservatively instead, raising stable accuracy in all three settings while leaving shift accuracy unchanged to within one point on two of them. Table 9 reports the per-setting false-switch rates. Table 9: Stable and shift steps improve together where the trajectory carries information. Decomposition of the shared-cache replay by whether the true state changed at that step (%). A policy that merely revises more often would buy shift accuracy at the cost of stable accuracy and a higher false-switch rate. On HumDial-En (human recordings with many true shifts), EmoUpdate improves shift accuracy on all three backbones while improving stable accuracy on two and preserving it on the third, with fewer spurious switches. On CREMA-D-Stream, whose transitions are uniform by construction and where the shrinkage estimator returns α=0α=0 (Eq. 6), it is conservative rather than aggressive: stable accuracy rises in all three settings without a gain on shift steps. Data Model Method Stable ↑ Shift ↑ False sw. ↓ Delay ↓ TF1 ↑ HumDial-En Qwen2-Audio ICI 40.00 67.43 71.43 0.07 88.89 C-HMM 40.00 66.86 68.57 0.07 87.57 EmoUpdate 54.29 86.86 60.00 0.04 92.31 Phi-4M ICI 88.57 80.00 17.14 0.01 95.68 C-HMM 88.57 80.00 17.14 0.01 95.68 EmoUpdate 97.14 84.57 8.57 0.01 96.21 MiniCPM-o 2.6 ICI 91.43 89.14 17.14 0.02 96.28 C-HMM 91.43 89.14 17.14 0.02 96.28 EmoUpdate 91.43 91.43 14.29 0.01 96.85 CREMA-D-Stream Qwen2-Audio ICI 74.92 74.59 24.26 0.13 57.52 C-HMM 78.36 69.67 16.07 0.14 63.35 EmoUpdate 78.36 69.67 16.07 0.14 63.35 Phi-4M ICI 20.16 17.21 31.64 0.32 22.03 C-HMM 20.66 17.21 20.49 0.25 17.71 EmoUpdate 20.66 17.21 20.49 0.25 17.71 MiniCPM-o 2.6 ICI 25.90 22.13 41.48 0.27 23.53 C-HMM 26.07 19.67 31.48 0.29 17.97 EmoUpdate 27.05 21.31 62.30 0.57 19.10 Figure 6: Contamination, and its repair, on every backbone. Writing the previous label into the prompt (red) costs up to 49 points of causal step accuracy relative to prior-blind inference (grey) across all eight settings; EmoUpdate (green) is highest in all eight. Same quantities as the step-accuracy columns of Table 2, drawn for comparison at a glance. Appendix G Compute and Fairness Audit The test-time contract is identical in every row of Table 10. Prompt development calls are a one-time cost for choosing s⋆s and are not repeated per policy or test seed. More precisely, all belief policies share the posterior cache produced by the selected grounding prompt. We do not claim that the unoptimized seed prompt and selected grounding prompt are the same. The fairness claim is therefore specific: deployment calls, trainable parameters, chunk boundaries, decoding, label extraction, and policy-candidate budgets are matched. Total development SpeechLM calls are not identical, because prompt selection is a component of EmoUpdate and costs the explicitly reported ten prompt candidates. Within any fixed perception source (seed or selected grounding), every policy comparison replays exactly the same cache. Table 10: Compute audit: zero extra model calls, zero trainable parameters. Data Model Test chunks Calls/chunk Test calls Prompt cand. Dev calls Policy cand. HumDial Qwen2-Audio 315 1 315 10 1,350 10 Qwen2.5-Omni 315 1 315 10 1,350 10 Phi-4M 315 1 315 10 1,350 10 MiniCPM-o 2.6 315 1 315 10 1,350 10 CREMA-D-Stream Qwen2-Audio 976 1 976 10 4,800 10 Qwen2.5-Omni 976 1 976 10 4,800 10 Phi-4M 976 1 976 10 4,800 10 MiniCPM-o 2.6 976 1 976 10 4,800 10 Appendix H Additional Intervention Results Figure 7: Where the prediction lands under each wrong injected prior (CREMA-D-Stream, Qwen2-Audio). The leading terracotta segment is the share captured by the injected label (identical to Figure 3), and the slate segments give the destinations of the remainder. Eligibility is the audit rule used throughout: a condition counts only when the injected prior differs from the prior-blind control prediction. A wrong happy prior is adopted 4.84.8% of the time and its mass lands mostly on neutral, whereas a wrong fearful prior is adopted 98.298.2% of the time. Table 11 reports all saved history-intervention summaries. CREMA-D-Stream uses six injected labels plus the control (seven calls per sample). HumDial uses seven injected labels plus the control (eight calls per sample). “Evidence wins” is the fraction of wrong-prior conditions where the control was correct and the exposed prediction remains correct. The Phi-4M accuracy increase under exposure is not evidence that history is harmless: its 43.67% flip rate and posterior JSD of 0.30 show high sensitivity, while the added label text acts as a regularizer for an otherwise weak control prompt. Figure 8: All ten grounding candidates under the same balanced injection. Dark bars state explicitly that the previous label must not be used; ⋆ is the deployed candidate, which never receives a prior at test time. No instruction falls below 59.7% wrong-prior pull, the three most emphatic are the worst, and pull rises with prior-blind accuracy (r=0.81r=0.81). The green rule marks what architectural removal achieves. Figure 9 repeats the diagnostic on HumDial-En for two further backbones. (a) Qwen2.5-Omni (b) Phi-4M Figure 9: HumDial-En previous-belief diagnostics under three prompt conditions. Phi-4M’s accuracy rises under exposure while its flip rate holds at 43.67%, so accuracy change alone is not a contamination measure. Table 11: Additional history-intervention summaries (%); negative drop means exposure helps. Data Model Condition N Ctrl. Exposed Drop Flip Wrong pull Evidence wins JSD HumDial Qwen2.5-Omni Grounded 35 97.14 95.10 2.04 2.04 0.48 97.55 0.03 Naive 35 97.14 92.65 4.49 4.49 1.90 95.10 0.03 Sticky 35 97.14 93.88 3.27 3.27 0.95 97.06 0.02 HumDial Phi-4M Grounded 35 45.71 84.49 −38.78-38.78 43.67 7.62 98.96 0.30 Naive 35 45.71 74.69 −28.98-28.98 43.67 18.57 96.88 0.29 Sticky 35 45.71 83.27 −37.55-37.55 43.67 11.43 100.00 0.29 CREMA-D-Stream Qwen2-Audio Prior-injected 120 72.50 30.42 42.08 65.69 71.00 23.68 0.48 Table 12 resolves the balanced grid by injected label. Table 12: Per-prior breakdown, CREMA-D-Stream Qwen2-Audio (%; JSD on 0–100). Injected belief Exposed accuracy ↑ Prior pull ↓ Posterior JSD ↓ neutral 49.17 39.13 19.78 happy 50.00 4.76 23.24 sad 23.33 90.32 49.41 angry 21.67 92.05 50.61 disgust 20.00 96.40 62.68 fearful 18.33 98.20 83.67 Is the asymmetry a label-surface artefact? Pull is not explained by label spelling or by the model’s own prior-blind base rate. Across the six labels the pull rate is uncorrelated with surface length (r=+0.020r=+0.020) and with the prior-blind base rate (r=+0.022r=+0.022). The clearest counterexample is internal: happy and angry have identical spelling length and comparable base rates (11.9% versus 23.2%) yet differ by a factor of 19 in pull (4.76% versus 92.05%); fearful has the lowest base rate (6.4%) and the highest pull. The pattern is consistent with a semantic rather than a tokenisation origin. Appendix I Measurement-Based Decontamination The intervention grid also supports repair when the perception prompt cannot be firewalled, as in legacy or third-party serving stacks. For every audited sample the grid stores the clean posterior q(∅)q^( ) and the contaminated posterior q(b)q^(b) under each injected belief b. We estimate the per-prior log-offset δb _b of Equation 7 on one stratified half of the grid and correct held-out contaminated posteriors as q^∝q(b)exp(δb) q q^(b) ( _b). The operator is closed-form, adds no SpeechLM calls, and needs only the system’s own previous label, which the deployed system itself wrote into the prompt. Results are averaged over the two folds. Figure 10: Measurement-based decontamination (CREMA-D-Stream, Qwen2-Audio; two-fold cross-validated). The operator recovers 56% of the accuracy damage and collapses wrong-prior pull from 71.0 to 12.2. Table 13 gives the per-fold numbers behind that figure. Table 13: Two-fold cross-validated decontamination, CREMA-D-Stream Qwen2-Audio (%). Quantity Contaminated Corrected Prior-blind control Exposed accuracy ↑ 30.42 53.89 72.50 Prediction flip rate ↓ 65.69 39.31 – Wrong-prior pull ↓ 71.00 12.17 – The correction recovers 56% of the accuracy damage (30.42→53.8930.42→ 53.89 against a 72.5072.50 control) and reduces the defining PBC symptom, wrong-prior pull, from 71.0071.00 to 12.1712.17 (Figure 10). Deployment scope. The operator is intended as a fallback for serving stacks where firewalling is unavailable, so the primary EmoUpdate deployment uses firewalling whenever prompt access is available. The cross-validated analysis above evaluates the correction on the CREMA-D-Stream/Qwen2-Audio intervention grid at the step level, conditioned on the injected label, and requires only that the deployment-time prior is observed, which holds whenever the system itself maintains the belief it exposes. How much intervention grid does the operator need? We sub-sample the training fold at fixed fractions, refit δb _b, and evaluate on the untouched held-out fold (three seeds × two folds). Table 14 shows the operator saturates almost immediately: 36 conditions (six chunks with all six priors injected) already recover exposed accuracy from 28.61 to 53.33, inside the seed-to-seed range of the 360-condition fit (53.89). A practical deployment therefore needs only a few dozen audited chunks, and refreshing is cheap. Table 14: Decontamination accuracy versus the size of the fitting grid (CREMA-D-Stream, Qwen2-Audio; held-out fold, three seeds × two folds). Uncorrected exposed accuracy is 28.61 and the prior-blind control is 72.50. Fitting conditions Exposed acc. ↑ Range Wrong-prior pull ↓ 36 53.33 45.28–62.50 12.00 72 54.54 47.50–58.06 14.44 108 54.72 50.28–58.06 13.44 144 54.44 51.11–57.22 13.89 180 54.49 51.39–58.06 12.94 216 54.54 50.83–57.22 12.33 288 54.58 51.11–58.06 11.56 360 53.89 52.78–55.00 12.17 Appendix J Asymmetry-Aware Transition Prior Table 15 reports the full comparison behind Section 4.6. Both variants replay identical posterior caches with the same pre-specified ten-candidate ρ grid; the asymmetric leave distribution is an add-one-smoothed first-order estimate from development gold sequences and introduces no additional tuned scalar. Table 15: Transition-prior variants (%, point estimates). All variants replay identical caches under the same ten-candidate budget. α is the shrinkage coefficient estimated from development transitions (Equation 6): it is 0 on CREMA-D-Stream, where the evidence-shrunk prior therefore coincides with the symmetric one by construction. Data Model Transition prior Final Step TF1 S-BAcc HumDial-En Qwen2-Audio (α=0.87α=0.87) Symmetric 88.57 69.84 63.84 53.43 Full asym. 84.76 80.95 78.99 72.57 Evidence-shrunk 86.67 82.22 81.54 70.29 + endpoint calib. 87.62 82.54 81.87 70.57 Qwen2.5-Omni (α=0.87α=0.87) Symmetric 91.43 90.48 90.49 95.14 Full asym. 90.48 91.43 92.49 96.00 Evidence-shrunk 90.48 91.43 92.49 96.00 + endpoint calib. 91.43 91.75 92.80 96.29 Phi-4M (α=0.87α=0.87) Symmetric 76.19 78.10 76.74 84.86 Full asym. 72.38 79.68 80.59 88.57 Evidence-shrunk 72.38 79.68 80.59 88.57 + endpoint calib. 80.00 82.22 84.55 90.86 MiniCPM-o 2.6 (α=0.87α=0.87) Symmetric 91.43 83.81 88.83 90.29 Full asym. 89.52 84.76 90.80 91.14 Evidence-shrunk 89.52 84.44 89.97 90.86 + endpoint calib. 91.43 85.08 91.12 91.43 CREMA-D-Stream Qwen2-Audio (α=0.00α=0.00) Symmetric 77.46 76.23 51.55 74.02 Full asym. 75.00 75.00 46.49 74.67 Evidence-shrunk 77.46 76.23 51.55 74.02 + endpoint calib. 77.46 76.23 51.55 74.02 Qwen2.5-Omni (α=0.00α=0.00) Symmetric 52.46 50.31 23.65 50.90 Full asym. 52.87 50.41 23.37 50.00 Evidence-shrunk 52.46 50.31 23.65 50.90 + endpoint calib. 67.21 54.00 22.89 53.85 Phi-4M (α=0.00α=0.00) Symmetric 20.90 19.88 4.43 18.93 Full asym. 20.49 19.88 4.76 18.93 Evidence-shrunk 20.90 19.88 4.43 18.93 + endpoint calib. 20.90 19.88 4.43 18.93 MiniCPM-o 2.6 (α=0.00α=0.00) Symmetric 26.23 24.59 5.24 23.36 Full asym. 26.64 24.49 4.96 22.95 Evidence-shrunk 26.23 24.59 5.24 23.36 + endpoint calib. 28.69 25.20 4.68 24.18 Finite-sample behaviour of α. Equation 6 is a point estimate, so we resample the development destination transitions with replacement (2,000 draws) and recompute α on each. On HumDial-En (n=75n=75 transitions) the coefficient is α=0.87α=0.87 with a 95% percentile interval [0.74,0.93][0.74,0.93], and no resample returns zero. On CREMA-D-Stream (n=60n=60) the point estimate is α=0.00α=0.00 and 60% of resamples also return exactly zero, though the interval [0.00,0.64][0.00,0.64] is wide: with sixty transitions the data cannot exclude moderate asymmetry, only fail to evidence it. This is the intended behaviour: the estimator abstains rather than asserting structure, but it is an abstention, not a proof of symmetry. Note α depends on the benchmark’s gold transitions alone, so it is shared by all four backbones. Appendix K Conditional Volatility Table 2 in the main paper reports the primary and conventional metrics. Table 16 adds the mechanism diagnostics: transition F1, inertia, conditional volatility, and revision delay. The inertia–volatility pair exposes how the causal filter balances persistence and revision: on HumDial-En EmoUpdate reduces inertia to under 3.1% for every backbone, while CREMA-D-Stream exhibits model-dependent trade-offs between inertia and volatility. Table 17 reports conditional volatility per method; it conditions on a previously correct stable belief and therefore has empty denominators for some conservative policies. Table 16: Mechanism metrics (%, point estimates; revision delay in chunks). All four are lower-better except TF1. Inertia is the fraction of true shifts the policy fails to follow; volatility is the fraction of stable steps it destroys; “Best ctl.” is the per-metric best controlled baseline. Data Model Method TF1 ↑ Inertia ↓ Vol. ↓ Delay ↓ HumDial-En Qwen2-Audio Best ctl. 56.49 17.71 0.00 0.000 EmoUpdate 81.87 2.86 53.57 0.038 Qwen2.5-Omni Best ctl. 87.68 3.47 9.52 0.013 EmoUpdate 92.80 2.50 0.00 0.012 Phi-4M Best ctl. 41.54 10.61 0.00 0.000 EmoUpdate 84.55 2.82 3.03 0.020 MiniCPM-o 2.6 Best ctl. 85.29 7.30 9.09 0.000 EmoUpdate 91.12 2.88 9.09 0.012 CREMA-D-Stream Qwen2-Audio Best ctl. 48.15 6.10 0.00 0.000 EmoUpdate 51.55 14.74 6.52 0.141 Qwen2.5-Omni Best ctl. 18.36 15.09 25.77 0.306 EmoUpdate 22.89 6.15 24.83 0.367 Phi-4M Best ctl. 4.44 52.17 0.00 0.000 EmoUpdate 4.43 61.54 16.24 0.250 MiniCPM-o 2.6 Best ctl. 4.84 75.00 0.83 0.091 EmoUpdate 4.68 56.25 53.47 0.567 Table 17: Conditional volatility (%, lower is better; point estimates). Dashes mark an empty conditional denominator, not a missing run. HumDial-En CREMA-D-Stream Method Qwen2-A. Qwen2.5 Phi-4M MiniCPM Qwen2-A. Qwen2.5 Phi-4M MiniCPM DHC – 9.52 0.00 9.09 0.00 65.38 35.06 0.83 ICI 100.00 26.67 100.00 23.33 15.42 30.08 0.00 22.63 CGR – 26.67 – 23.33 14.52 30.08 0.00 22.63 MGR – 26.67 100.00 23.33 13.82 30.08 0.00 22.46 C-HMM 50.00 30.00 100.00 24.14 13.58 30.89 0.00 23.19 IJSR 0.00 23.33 100.00 10.34 10.38 25.77 0.00 20.59 AJSR – 16.67 – 20.69 10.44 28.02 0.00 18.98 EmoUpdate 53.57 0.00 3.03 9.09 6.52 24.83 16.24 53.47 Appendix L Published-Method Adaptation DVL-CER (Zha et al. 2025) uses learned conversational representations rather than frozen SpeechLM posteriors. We therefore implement a causal posterior-GRU adaptation following the published formulation and report it separately from the compute-matched controlled envelope, since its randomized adaptation runs are not the same uncertainty object as our episode-bootstrap replicates. Table 19 gives the complete comparison. Table 18: Published DVL-CER adaptation versus EmoUpdate (%). Positive Δ favors EmoUpdate; uncertainty estimates are reported in Table 19. TF1 ↑ S-BAcc ↑ Data Model DVL EmoUpdate Δ DVL EmoUpdate Δ CREMA-D-Stream Qwen2-Audio 19.83 51.55 +31.72 48.25 74.02 +25.77 Qwen2.5-Omni 10.14 22.89 +12.75 32.05 53.85 +21.80 Phi-4M 2.53 4.43 +1.90 17.10 18.93 +1.83 MiniCPM-o 2.6 3.99 4.68 +0.69 19.33 24.18 +4.85 HumDial-En Qwen2-Audio 13.23 81.87 +68.64 24.86 70.57 +45.71 Qwen2.5-Omni 18.01 92.80 +74.79 35.71 96.29 +60.58 Phi-4M 16.53 84.55 +68.02 31.81 90.86 +59.05 MiniCPM-o 2.6 23.53 91.12 +67.59 24.89 91.43 +66.54 Table 19: Paper-derived DVL-CER adaptation versus EmoUpdate (%). DVL-CER reports mean± deviation over three randomized adaptation runs; EmoUpdate reports the full-set point estimate± -bootstrap standard error (B=1000B=1000). Data Model Method Step TF1 S-BAcc CREMA-D-Stream Qwen2-Audio DVL-CER adapt. 51.57±6.3351.57± 6.33 19.83±5.5119.83± 5.51 48.25±5.2948.25± 5.29 EmoUpdate 76.23±2.0676.23± 2.06 51.55±3.7151.55± 3.71 74.02±2.5874.02± 2.58 Qwen2.5-Omni DVL-CER adapt. 35.38±8.7035.38± 8.70 10.14±4.4010.14± 4.40 32.05±8.5832.05± 8.58 EmoUpdate 54.00±2.0854.00± 2.08 22.89±2.4422.89± 2.44 53.85±2.5153.85± 2.51 Phi-4M DVL-CER adapt. 17.49±0.9017.49± 0.90 2.53±0.002.53± 0.00 17.10±0.6717.10± 0.67 EmoUpdate 19.88±1.9919.88± 1.99 4.43±1.744.43± 1.74 18.93±2.1818.93± 2.18 MiniCPM-o 2.6 DVL-CER adapt. 22.17±3.5122.17± 3.51 3.99±2.313.99± 2.31 19.33±4.3519.33± 4.35 EmoUpdate 25.20±1.6525.20± 1.65 4.68±1.294.68± 1.29 24.18±1.9724.18± 1.97 HumDial-En Qwen2-Audio DVL-CER adapt. 21.38±8.8921.38± 8.89 13.23±10.8213.23± 10.82 24.86±31.7024.86± 31.70 EmoUpdate 82.54±2.0282.54± 2.02 81.87±2.8181.87± 2.81 70.57±4.1670.57± 4.16 Qwen2.5-Omni DVL-CER adapt. 26.46±12.5826.46± 12.58 18.01±16.7218.01± 16.72 35.71±29.3035.71± 29.30 EmoUpdate 91.75±1.5191.75± 1.51 92.80±1.8692.80± 1.86 96.29±0.9596.29± 0.95 Phi-4M DVL-CER adapt. 24.23±12.8624.23± 12.86 16.53±16.6616.53± 16.66 31.81±30.6531.81± 30.65 EmoUpdate 82.22±2.2582.22± 2.25 84.55±2.6984.55± 2.69 90.86±2.2290.86± 2.22 MiniCPM-o 2.6 DVL-CER adapt. 20.48±12.9520.48± 12.95 23.53±12.7123.53± 12.71 24.89±30.8924.89± 30.89 EmoUpdate 85.08±1.9785.08± 1.97 91.12±2.0491.12± 2.04 91.43±2.8391.43± 2.83 Appendix M Cross-Benchmark Component Analysis Table 20 expands Table 3 to every reported metric, backbone, and benchmark, under the same bootstrap protocol. Table 20: Complete component analysis across four backbones and both benchmarks (%, point estimates). “Neither”: seed prompt with chunkwise updates; “Policy only”: seed prompt with the causal filter; “Ground only”: prior-blind prompt with chunkwise updates; “Both”: the deployed system. Volatility is lower-better; a dash marks an empty conditional denominator. Data Model Components Final Step Step-MaF1 TF1 S-BAcc Vol. HumDial-En Qwen2-Audio Neither 89.52 60.95 68.28 50.75 29.14 100.00 Policy only 89.52 66.35 72.97 58.55 37.43 83.33 Ground only 87.62 70.16 71.40 63.33 53.71 61.54 Both (EmoUpdate) 87.62 82.54 82.31 81.87 70.57 53.57 Qwen2.5-Omni Neither 89.52 86.03 86.14 86.61 81.43 26.67 Policy only 89.52 86.35 86.26 87.54 86.29 16.13 Ground only 91.43 90.16 89.55 90.86 91.43 9.38 Both (EmoUpdate) 91.43 91.75 90.66 92.80 96.29 0.00 Phi-4M Neither 67.62 43.81 41.82 35.80 21.14 100.00 Policy only 71.43 49.52 48.37 44.12 28.57 100.00 Ground only 76.19 78.73 74.88 77.81 84.29 6.45 Both (EmoUpdate) 80.00 82.22 78.54 84.55 90.86 3.03 MiniCPM-o 2.6 Neither 93.33 84.13 84.67 84.57 79.71 23.33 Policy only 93.33 81.27 82.18 81.31 80.57 12.50 Ground only 91.43 83.81 82.10 88.83 90.29 6.45 Both (EmoUpdate) 91.43 85.08 82.95 91.12 91.43 9.09 CREMA-D-Stream Qwen2-Audio Neither 70.08 69.57 68.41 39.80 70.33 15.42 Policy only 70.08 69.98 68.45 41.21 68.69 11.53 Ground only 74.59 74.69 73.44 45.91 74.75 13.94 Both (EmoUpdate) 77.46 76.23 74.58 51.55 74.02 6.52 Qwen2.5-Omni Neither 48.77 43.85 42.21 17.65 42.30 30.08 Policy only 57.79 46.41 45.40 17.67 45.00 29.73 Ground only 52.87 50.41 47.23 23.61 50.66 24.75 Both (EmoUpdate) 67.21 54.00 52.45 22.89 53.85 24.83 Phi-4M Neither 15.98 16.39 4.70 – 15.98 0.00 Policy only 18.03 16.91 9.17 – 16.39 33.00 Ground only 20.90 19.57 11.20 4.52 18.69 29.06 Both (EmoUpdate) 20.90 19.88 10.85 4.43 18.93 16.24 MiniCPM-o 2.6 Neither 22.13 22.23 13.75 4.12 20.08 22.63 Policy only 28.69 23.77 18.66 2.65 21.31 38.41 Ground only 25.41 24.59 17.18 6.12 24.02 33.79 Both (EmoUpdate) 28.69 25.20 22.61 4.68 24.18 53.47 The decomposition is consistent with the two-stage firewalled inference path: prior-blind grounding supplies the largest change in observation quality, while the same causal filter yields its strongest gains once that observation is usable. Metric-specific trade-offs, including the CREMA-D-Stream settings where the filter exchanges volatility or TF1 for balanced accuracy, are reported in Table 20. Appendix N Selection-Budget and Parameter Sensitivity Table 21 replays prefixes of the pre-specified joint grid. Two of the eight settings lock their final choice by three candidates; the others keep refining up to the full budget. The selection is nevertheless not delicate: as Figure 5 shows, the metric is flat across a wide central band of ρ, so the candidates a smaller budget would have chosen sit on the same plateau. Table 21: Selection-budget sensitivity. Each cell is the candidate (ρ/endpoint mode) chosen when only the first k of the ten pre-specified joint candidates are available; “def” is the acoustic endpoint and “LS” the label-shift endpoint. A check mark means budgets 3, 5, and 10 agree. Data Model k=1k=1 k=3k=3 k=5k=5 k=10k=10 Stable HumDial-En Qwen2-Audio 0.1/def 0.1/def 0.1/def 0.1/def ✓ Qwen2.5-Omni 0.1/def 0.1/def 0.4/def 0.4/def Phi-4M 0.1/def 0.2/def 0.4/def 0.5/LS MiniCPM-o 2.6 0.1/def 0.1/LS 0.1/LS 0.1/LS ✓ CREMA-D-Stream Qwen2-Audio 0.1/def 0.2/def 0.4/def 0.4/def Qwen2.5-Omni 0.1/def 0.1/LS 0.1/LS 0.1/LS ✓ Phi-4M 0.1/def 0.2/def 0.2/def 0.2/def ✓ MiniCPM-o 2.6 0.1/def 0.1/LS 0.1/LS 0.1/LS ✓ That replay covers both benchmarks and all evaluated backbones. Performance degrades most visibly near ρ=1ρ=1, where the filter approaches an always-preserve policy and true revisions are suppressed; elsewhere the curves are flat, which is why the budget prefixes in Table 21 land on comparable operating points even when they select different candidates. Appendix O Additional Diagnostic Analyses Table 22 collects every pre-specified diagnostic case in one place. Table 22: Pre-specified diagnostic cases across model–benchmark settings (gaps in points). Setting Metric or diagnostic Observation Implication HumDial, Qwen2-Audio Final accuracy −1.90-1.90 vs. ICI Final-only ranking hides trajectory gains CREMA-D-Stream, Phi-4M TF1 −0.02-0.02 vs. DHC Low-information acoustic posterior HumDial, Phi-4M Prior exposure Accuracy +38.78+38.78, flip 43.6743.67 Text may regularize yet still alter perception HumDial stable steps Volatility Some denominators empty Report “–” rather than impute a score The first row reinforces the need for trajectory metrics: EmoUpdate improves step accuracy, TF1, and S-BAcc by 21.59, 31.12, and 41.43 points against the same baseline in the same setting. The Phi-4M CREMA-D-Stream row illustrates the low-information regime: that current-audio posterior is heavily concentrated on a small subset of labels, so temporal structure has limited acoustic evidence to reweight. The HumDial intervention shows why accuracy change alone is not a PBC measure: sensitivity is established by holding audio fixed and measuring output change, regardless of whether the incidental change helps or hurts a particular label mix. Appendix P Reproducibility and Responsible Use Every run stores episode manifests, posterior caches, selected-candidate locks, configuration hashes, per-episode predictions, metric summaries, and plot-source tables. Table 23 maps each claim to its reproducibility artifact. Selection locks include candidate order, development metrics, selected parameters, prompt hashes, data hashes, and cache hashes, preventing post-evaluation candidate changes. Table 23: Reproducibility artifact inventory. Artifact Contents Episode manifests Ordered chunks, labels, split keys, source-audio hashes Perception caches One dense label posterior per frozen-model chunk call Selection locks Candidate budget, development ranking, chosen prompt/policy Policy predictions Per-step prior, action, posterior, and predicted state Bootstrap summaries Point estimates with B=1000B=1000 bootstrap standard errors Intervention records Control and every injected-prior posterior Plot-source CSVs Every point and bar shown in the paper figures Paper figure script Deterministic vector/PDF figure generation End-to-end reproduction order. A clean reproduction follows six locked stages: (1) verify source-audio and episode-manifest hashes; (2) run each frozen SpeechLM once per development chunk for every prompt candidate; (3) select the prompt using only development metrics and write the selection lock; (4) generate one evaluation posterior cache with the selected prompt; (5) select the causal policy on cached development posteriors and replay the locked policy on evaluation posteriors; and (6) compute full-set metrics, B=1000B=1000 episode-bootstrap standard errors, and figures from the saved predictions. No evaluation label is read during prompt or policy selection, and bootstrap seeds alter only episode weights, never model outputs or selected hyperparameters. The source data retain their original licenses; constructed trajectories contain no generated emotion labels. Emotion recognition is culturally and contextually uncertain and should not be used as a clinical diagnosis, a hiring signal, or a basis for consequential surveillance. PBC also shows that system state can amplify an erroneous earlier label; applications should expose uncertainty and permit correction rather than treating filtered beliefs as facts.