Some Emotions Run Deeper: Layer-wise Probing and Causal Intervention in Large Language Models
Tian Fang, Gaël Guibon, Davide Buscaldi
cs.CL · cs.AI
Quick verdict
Claims emotion-probing depth in LLMs is dataset-dependent; multi-corpus probing plus interventions broadly support this, with modest, family-dependent effects.
Summary
The paper asks where emotion information becomes accessible in decoder-only LLMs and whether that depth is model- or corpus-driven. The authors probe eight open-weight models (Llama-3.2/3.1, Qwen-3.5, Granite-4.1; 1B-9B) on three English emotion corpora restricted to fear, joy, anger, and sadness: Emotion/CARER tweets, GoEmotions Reddit comments, and ISEAR autobiographical narratives. Using frozen concat(mean,max,min) pooling with logistic-regression probes, the best decodable depth shifts from about 0.066 relative depth on Emotion to 0.219 on GoEmotions and 0.590 on ISEAR, an ordering that survives label-by-length-bin matching. Online forward interventions scaling probe-selected bands drop test accuracy 26.6 points versus baseline, exceeding same-width unrelated bands by 6.1 points (q<0.001), though specificity is non-significant for Granite and weak offline (-0.8, n.s.). Selected bands transfer across datasets and emotions, and early-exit classifiers read from selected bands beat full-depth exits by 6.9 points (256 settings, q<0.001), plus 15.8 and 21.5 points over frozen RoBERTa and DeBERTa features. Limitations: interventions and early exit cover only two of the three corpora, English-only data, frozen models with linear readout, and effects that vary by dataset and model family.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 2
● low 3
4 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- The layer at which emotion becomes linearly decodable in decoder-only LLMs shifts systematically with the text source, from ~0.066 relative depth on Emotion/CARER to ~0.590 on ISEAR, and this ordering persists after label-by-length-bin matching.
- Online forward-pass scaling of probe-selected layer bands reduces test accuracy 5-6 points more than same-width unrelated bands (q<0.01) across 768 matched settings.
- Selected bands transfer across datasets and emotion categories, indicating partially shared affective information rather than strictly per-emotion substrates.
- Early-exit classifiers reading probe-selected bands outperform full-depth exits by 6.9 percentage points on average across 256 settings, and frozen RoBERTa/DeBERTa features by 15.8/21.5 points.
Strengths
Cross-corpus design directly tests the central question
Unlike prior single-corpus probing (Di Palma et al. 2025), Section 4 spans 3 datasets x 8 models (24 dataset-model pairs), showing best decodable depth shifts from 0.066 (Emotion) to 0.219 (GoEmotions) to 0.590 (ISEAR) relative depth — turning 'emotion lives in early layers' into an empirical, source-dependent question.
Confound controls are built in, not bolted on
Label-by-length-bin matched subsets preserve the depth ordering (0.131/0.288/0.642, Section 4); corpus-cue diagnostics (Table 4) quantify cue profiles on matched subsets; OvR vs 4-class probe consistency is checked (r=0.85, 87.5% in-range, Section 4 and Fig. 11).
Causal claims carry matched controls and multiple-comparison hygiene
Unusually honest negative results
Weaknesses
Causal and early-exit coverage omits the intermediate corpus
Specificity is modest and sits atop a large generic disruption
Early-exit framing overstates the result
Amplification asymmetry is unexplained
Selection stability and baseline strength are unquantified
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
The paper asks whether the depth at which emotion becomes linearly decodable in decoder-only LLMs is a model-universal fact or a property of the text source. It probes eight open-weight models — Llama-3.2-1B/3B-Instruct, Llama-3.1-8B-Instruct, Qwen-3.5-2B/4B/9B, Granite-4.1-3B/8B (Section 3) — on three English corpora restricted to fear/joy/anger/sadness: Emotion/CARER (tweets), GoEmotions (Reddit), and ISEAR (autobiographical narratives). Four claims are made (Section 1): (i) best probing depth shifts systematically across corpora; (ii) online interventions on probe-selected bands are 5-6 points more disruptive than same-width unrelated bands; (iii) bands transfer across datasets and emotions; (iv) early exits from selected bands beat full-depth readout by 6.9 points. Claim (i) is well supported; (ii)-(iv) are supported only on Emotion and ISEAR (Sections 5.1 and 6) and carry caveats the paper itself flags.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 4 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
Do emotions live in the early layers of LLMs?
Not universally, according to this paper. The best layer for decoding emotion depends on the text source: near the input for tweets (relative depth ~0.07), intermediate for Reddit comments (~0.22), and past mid-depth for autobiographical narratives (~0.59). The ordering persists even when texts are matched by label and length.
Is this finding causal or just correlational probing?
The authors go beyond probing by scaling activations during the forward pass. Probe-selected bands hurt downstream emotion accuracy about 6 points more than random same-size bands (q<0.001 over 768 settings), supporting causal engagement — though the effect is absent for Granite models, weak for some band widths, and does not appear when features are edited offline after extraction.
Can you stop an LLM early for emotion classification?
In their setup, a linear classifier reading a probe-selected intermediate band beats one reading the final layers by about 6.9 accuracy points on average, and beats frozen RoBERTa/DeBERTa features by roughly 16-21 points. This is a readout-location result under frozen models, not a general claim that early exit improves accuracy.
Which datasets and models were used in this study?
Three English corpora — Emotion/CARER (tweets), GoEmotions (Reddit comments, 4-emotion single-label subset), and ISEAR (autobiographical narratives) — restricted to fear, joy, anger, sadness; and eight open-weight decoder-only models from Llama (1B-8B), Qwen-3.5 (2B-9B), and Granite-4.1 (3B-8B).
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figure curves (Figs. 2, 3, 6-12) appear in the extraction only as axis labels and legends; figure-level patterns are assessed from captions and in-text numbers, not the plotted data.
- No code repository URL appears in the provided text; code release is implied ('released diagnostic code' in App. A; supplementary CSVs) but its scope and location cannot be verified from the text.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.