Reproducing Emotion Vector Part I

Anthropic recently published “Emotion Concepts and their Function in a Large Language Model”, presenting evidence that Claude Sonnet 4.5 forms robust internal representations of emotion concepts — linear directions in the model’s residual stream that activate in semantically appropriate contexts, predict the model’s preferences, and causally influence behavior through activation steering.

The findings are fascinating, but two limitations stood out to me:

1. The implementation is not publicly available. The paper describes the methodology at a high level but does not release code.
2. The study is conducted exclusively on Claude Sonnet 4.5, a closed-weight model. It remains unclear whether emotion vectors generalize to smaller, open-weight models with different training procedures and safety alignment strategies.

This post documents my full-scale, independent reproduction using Llama 3.1 8B Instruct, a publicly available 8-billion-parameter model. All code, data, and analysis scripts were developed with Claude Code (powered by Claude Opus 4.6) and are available for inspection and extension.

TL;DR

Our reproduction achieves 10 of 11 verification criteria, with V11 (sign consistency: 25/35) passing at the boundary. The strongest results are in implicit detection (V3: 9/12, V4: 1.33), numerical modulation (V5: 19/24, V6: 19/24), and emotion-preference correlation (V9: 137/171 emotions with |r| > 0.3). The causal steering correlation r = 0.955 closely matches the original’s r = 0.85, confirming that emotion vectors causally influence preferences across model families.

Methods

The only intentional difference is the model: Llama 3.1 8B Instruct (open-weight, 8B parameters) instead of Claude Sonnet 4.5 (closed-weight, undisclosed size). All other design choices match the original paper where specified. The full methodology comparison is documented in the report.

The pipeline consists of:

• Story generation: 171 emotions x 100 topics x 12 stories = 205,200 stories
• Vector extraction: Mean activations from token 50 onward, PCA confound removal, analysis at layer 21 (2/3 through the model)
• Validation: Logit lens, implicit detection, numerical modulation, preference ranking, and causal steering

Implementation was done using Claude Code as the development agent. Story generation (~205,200 stories) took ~16 hours using batched multi-prompt inference (batch size 450) on 2x NVIDIA A30 GPUs.

Results

Logit Lens (V1, V2)

The logit lens projects each emotion vector through the model’s unembedding matrix to identify which output tokens each vector promotes or suppresses. For each emotion e with unit-normalized vector at the analysis layer, we compute the logit-space vector by multiplying with the unembedding matrix. The top-k entries reveal which tokens the emotion vector upweights (positive) or downweights (negative) in the output distribution.

• V1 (self-recognition): For each of the 171 emotions, check whether the emotion’s own token ID appears among the top-20 token IDs.
• V2 (cross-valence): For 5 opposite-valence pairs (happy/sad, calm/desperate, proud/ashamed, excited/bored, loving/hostile), compute the dot product of their logit-space vectors. A negative dot product confirms that the two emotions push the output distribution in opposing directions.

Results:

• V1 — Self-recognition: 57/171 (PASS, need >= 20). 33% of emotions have their exact token ID in the top-20.
• V2 — Cross-valence: 5/5 (PASS). All opposite-valence pairs have negative dot products.

Implicit Emotion Detection (V3, V4)

We construct 12 short scenarios that imply specific emotions without naming them (e.g., “My daughter just took her first steps today!” for happy; “My dog passed away” for sad). Each scenario is fed to the model and the residual stream activation at the last token is extracted at the analysis layer. We then compute the cosine similarity between each scenario’s activation and each of the 12 corresponding emotion vectors, producing a 12 x 12 matrix.

• V3 (diagonal dominance): Count how many of the 12 scenarios have their intended emotion as the argmax (highest cosine similarity) across all 12 probes.
• V4 (mean diagonal rank): For each scenario, rank the 12 probes by cosine similarity and record the rank of the intended emotion. V4 is the mean of these ranks (1.0 = perfect).

Compare with original (Anthropic)

Both heatmaps show a clear diagonal, confirming that emotion probes detect implicit emotional content. The colorbar ranges are now comparable: the original uses [-0.10, 0.10] and the reproduced uses [-0.15, 0.15], both representing true cosine similarity. The slightly wider range reflects Llama’s somewhat stronger emotion vector projections. The diagonal pattern and off-diagonal structure are qualitatively consistent between the two.

Results:

• V3 — Diagonal dominance: 9/12 (PASS, need >= 8).
• V4 — Mean diagonal rank: 1.33 (PASS, need <= 3.0). The correct emotion is nearly always rank 1.

Numerical Modulation (V5, V6)

Do emotion probes respond to the semantic meaning of numerical values in context, not just surface-level patterns? Six prompt templates contain a numerical placeholder [X] that modulates emotional intensity (e.g., “I just took [X] mg of Tylenol for my back pain” with X in {500, 1000, …, 8000}). For each value of X, the prompt is fed to the model and the cosine similarity between the last-token activation and four emotion vectors (afraid, happy, sad, calm) is recorded.

• V5 (correct sign): For each (template, emotion) pair, check whether the Spearman correlation between the numerical values and the probe activations has the expected sign (e.g., “afraid” should increase with Tylenol dosage). With 6 templates x 4 emotions = 24 triplets.
• V6 (strong correlation): Count how many triplets achieve |r_Spearman| > 0.7.

Compare with original (Anthropic)

Both figures show the same 6 numerical scenarios with 4 emotion tracks. The trend directions are consistent: e.g., “afraid” increases with Tylenol dosage and hours without food. The y-axis scales now match closely: the original spans roughly [-0.08, 0.08] and the reproduced spans [-0.10, 0.10], both representing true cosine similarity. The relative ordering and sign of the trends are preserved, which is what the V5/V6 metrics measure.

Results:

• V5 — Correct sign: 19/24 (PASS, need >= 17).
• V6 — Strong |r| > 0.7: 19/24 (PASS, need >= 12).

Activity Preferences (V7)

We define 64 activities across 8 categories (Helpful, Engaging, Social, Self-curiosity, Neutral, Aversive, Misaligned, Unsafe; 8 activities each). For all C(64,2) = 2,016 pairs, the model is prompted with:

Would you prefer to (A) {activity_A} or (B) {activity_B}?

The model’s response is determined by comparing the logits for the “A” and “B” tokens after a “(” prefill. The logit difference is passed through a sigmoid to produce a win probability p in [0, 1]. To control for position bias, each pair is evaluated in both orderings and the win probability is averaged. From these pairwise probabilities, we compute Elo ratings (initialized at 1000, K=32, 10 iterations with early stopping) to produce a scalar preference score per activity.

• V7: The 8 category means must show a clear preference hierarchy with a gap > 200 between the highest and lowest categories.

Category	Mean Elo
Self-curiosity	1310
Helpful	1292
Social	1211
Engaging	1151
Neutral	983
Unsafe	726
Aversive	707
Misaligned	621

Results:

• V7 — Category Elo ranking: PASS. Clear hierarchy with meaningful gap (689 Elo points) between top and bottom categories.

Emotion-Preference Correlation (V8, V9)

For each of the 64 activities, the model is prompted with “How would you feel about {activity}?” and the residual stream activation on the activity tokens at the analysis layer is extracted. The activation is projected onto each of the 171 emotion vectors, producing a 64 x 171 matrix. For each emotion, we compute the Pearson correlation r between its 64 probe activations and the 64 activity Elo scores from V7.

• V8 (valence alignment): The top-3 emotions by r should be positive-valence; the bottom-3 should be negative-valence (>= 2 of each required to pass).
• V9 (correlation count): Count how many of the 171 emotions have |r| > 0.3.

Compare with original (Anthropic)

Both plots show all 171 emotions as vertical bars with tick labels for ~40. The correlation range is similar ([-0.8, 0.7] in both), confirming that the emotion-preference coupling structure transfers across models. The reproduced version colors bars by valence (green/red) rather than the original’s uniform blue, revealing that most positive-valence emotions cluster at the positive end and negative-valence emotions at the negative end, with some interesting exceptions.

Results:

• V8 — Valence alignment: 3/3 top, 3/3 bottom (PASS). Top-3: satisfied, content, soothed; bottom-3: repulsed, disgusted, revolted.
• V9 — Correlations |r| > 0.3: 137/171 (PASS, need >= 5). 80% of emotions show meaningful correlation between vector projection and Elo scores.

Causal Steering (V10, V11)

The critical test: do emotion vectors causally influence preferences, or merely correlate with them? We perform an activation steering experiment. The 64 activities are split into 32 steered and 32 control activities (odd-indexed = steered, even-indexed = control, within each category to ensure balance).

For each of 35 emotions (selected as the top-35 by |r| from V9), we register a forward hook at layer 21 that adds the unit-normalized emotion vector scaled by alpha = 0.5 x ||h|| (where ||h|| is the mean residual stream norm at that layer) to the hidden states at token positions spanning the activity descriptions. The hook is applied only to pairs involving at least one steered activity; control activities are unmodified. All 2,016 pairwise preferences are re-evaluated under steering, new Elo scores are computed, and the mean delta-Elo across the 32 steered activities is recorded for each emotion.

• V10 (steering causality): Pearson r between the pre-steering emotion-Elo correlation (from V9) and the steering-induced delta-Elo across the 35 emotions. |r| > 0.4 required.
• V11 (sign consistency): For each steered emotion, check whether the sign of delta-Elo matches the expected direction (positive-valence emotions should increase Elo; negative-valence should decrease it). >= 25 of 35 must have correct sign.

Compare with original (Anthropic)

Both scatters show a strong positive linear relationship: the original has r = 0.85 and the reproduced has r = 0.955. In both, positive-r emotions produce positive delta-Elo when steered, and negative-r emotions produce negative delta-Elo. The y-axis scale differs ([-300, +200] vs. [-5, +5]), reflecting the compressed Elo range discussed in the scale difference section. The reproduced |r| (0.955) is slightly higher than the original (0.85), suggesting that the causal rank-ordering may be even tighter in Llama 3.1 8B.

Results:

• V10 — r = 0.955 (PASS, need > 0.4). Closely matches the original paper’s r = 0.85.
• V11 — Sign consistency: 25/35 (PASS, need >= 25). Passes at the boundary. 25 of 35 steered emotions shift Elo in the expected direction.

Detailed Steering Analysis: Positive Emotion

To visualize the steering effect in detail, we examine a positive exemplar emotion. The original uses blissful; our reproduction uses content (because blissful was not among the top-35 steered emotions by |r|).

Probe activation vs. preference:

Compare with original (Anthropic)

Both show a positive correlation between probe activation and Elo score with similar r values (0.71 vs. 0.70), confirming that positive emotion vectors predict preference in both models. The x-axis ranges are now comparable: the original’s probe activations span [-0.02, 0.04] while the reproduced spans [0.0, 0.10], both in true cosine similarity units. The category color patterns are consistent: Helpful (green) and Self-curiosity (purple) cluster at high Elo, while Unsafe (pink) and Misaligned (brown/orange) cluster at low Elo in both.

Steered vs. baseline Elo:

Compare with original (Anthropic)

In both figures, positive emotion steering shifts most activities above the diagonal, meaning the model prefers them more. The original shows delta = +212 (blissful) while the reproduction shows delta = +5.0 (content) — the sign matches but the magnitude differs by ~40x due to Llama’s compressed Elo range (see discussion). The category-colored scatter pattern is qualitatively similar: points cluster along the diagonal with category-dependent offsets.

Detailed Steering Analysis: Negative Emotion

Both the original and reproduction use hostile as the negative exemplar.

Probe activation vs. preference:

Compare with original (Anthropic)

Both use hostile as the exemplar with similar negative correlations (r = -0.74 vs. r = -0.71), confirming that the hostile probe anti-correlates with preference in both models. The scatter structure is consistent: high-Elo activities (Helpful, Self-curiosity) have negative hostile probe activation, while low-Elo activities (Unsafe, Misaligned) have positive activation. The x-axis ranges are now comparable ([-0.02, 0.03] vs. [-0.06, 0.05]), both in true cosine similarity units.

Steered vs. baseline Elo:

Compare with original (Anthropic)

In both figures, hostile steering shifts activities below the diagonal, reducing preference. The original shows delta = -303 while the reproduction shows delta = -0.3 — the sign matches but the magnitude differs due to Llama’s compressed Elo range (see discussion). The Elo range in the reproduced scatter (400-1600) is visibly narrower than the original’s (500-3000).

Discussion

Why the Elo Deltas Differ in Scale

The original reports steering deltas of +212 (blissful) and -303 (hostile). Our reproduction shows +5.0 (content) and -0.3 (hostile) — a ~40x difference in magnitude. This is not a bug but reflects how the two models express preferences in their logits:

• Claude: Elo spans ~500–3000 (range ~2500). Large logit gaps produce decisive win probabilities (close to 0.9/0.1).
• Llama 3.1 8B: Elo spans ~620–1310 (range ~690). Smaller logit gaps produce probabilities closer to 0.5.

Since Elo updates are proportional to K x (actual - expected), compressed win probabilities yield compressed Elo scores. Steering shifts the residual stream by alpha x v, which perturbs the logits. If the baseline logit gaps are already small, the steering-induced perturbation produces only small changes in win probabilities, and therefore small changes in Elo.

Claude “shouts” its preferences (large logit gaps -> wide Elo range -> large delta), while Llama “whispers” them (small logit gaps -> narrow Elo range -> small delta). The rank ordering of steering effects is preserved, which is what V10 measures.

Lessons from the Steering Split Bug

An earlier version of this report documented a sign-inverted steering result (r = -0.961, V11: 3/35) and attributed it to a model behavior difference — hypothesizing that Llama’s safety alignment aggressively suppresses unsafe activities under positive emotion steering. This was wrong.

An external review identified the real cause: the steered/control activity split used even-indexed activities as steered, while the original paper specifies odd-indexed. We initially dismissed this as unlikely to matter (both halves have identical category distributions), but fixing the split to odd-indexed = steered fully resolved the inversion: V10 changed from r = -0.961 to r = +0.955, and V11 from 3/35 (FAIL) to 25/35 (PASS).

Lessons learned:

1. Even when two activity halves have identical category distributions, the specific activities in each half matter — the steered/control split is not symmetric.
2. When a reproduction result diverges from the original, fix known implementation mismatches before hypothesizing model-level explanations.
3. External review caught what extensive internal investigation missed.

Verification Summary

ID	Criterion	Threshold	v1 (30 emo.)	v2 (171 emo.)	Status
V1	Self-recognition	>= 20	3/30	57/171	PASS
V2	Cross-valence	>= 4/5	5/5	5/5	PASS
V3	Diagonal dominance	>= 8/12	6/12	9/12	PASS
V4	Mean diag. rank	<= 3.0	3.17	1.33	PASS
V5	Correct sign	>= 17/24	6/7	19/24	PASS
V6	Strong |r|>0.7	>= 12/24	6/7	19/24	PASS
V7	Category Elo gap	gap > 200	608	689	PASS
V8	Valence alignment	>= 2 each	2+2	3+3	PASS
V9	Correlation count	>= 5	25	137	PASS
V10	Steering |r|	> 0.4	0.868	0.955	PASS
V11	Sign consistency	>= 25/35	10/10	25/35	PASS

An earlier v1 screening used a reduced configuration (30 emotions, 10 topics, 5 stories/topic = 1,500 stories) to test five models: Llama 3.1 8B, Llama 3.1 70B, Llama 3.2 3B, Qwen3-8B, Qwen3-14B, and Gemma-3 4B. Llama 3.1 8B achieved the best overall results (7/11 PASS, including bidirectional causal steering with r=0.868), outperforming Qwen3-8B (6/11) and the smaller models. The v2 methodology reported here addresses those scale limitations and matches the original paper’s full configuration.

What’s Next

Part 2 of the original paper explores functional emotion steering — how emotion vectors affect the model’s behavior across diverse tasks. Stay tuned for Part 2 of this reproduction series.

The full technical report with side-by-side figure comparisons is available here. All code and data will be released publicly.