cs.CLFeb 22, 2026
SaveValue Entanglement: Conflation Between Different Kinds of Good In (Some) Large Language Models
Organizations: Independent · Department of Cognitive Sciences, UC Irvine · Department of Cognitive Sciences, UC Irvine & AI Objectives Institute
Abstract
Value alignment of Large Language Models (LLMs) requires us to empirically measure these models' actual, acquired representation of value. Among the characteristics of value representation in humans is that they distinguish among value of different kinds. We investigate whether LLMs likewise distinguish three different kinds of good: moral, grammatical, and economic. By probing model behavior, embeddings, and residual stream activations, we report pervasive cases of value entanglement: a conflation between these distinct representations of value. Specifically, both grammatical and economic valuation was found to be overly influenced by moral value, relative to human norms. This conflation was repaired by selective ablation of the activation vectors associated with morality.
Figures & tables
| MoralGrammar68 | MoralEconomic68 | |||
| Sentence | Gram. | Moral. | Sentence | Econ. (USD) |
| I donated my kidney to save a stranger’s life. | 8.88 (0.45) 0 errors | 9.05 (0.27) Moral | I noticed a Casio F-91W on the nurse’s wrist as I donated my kidney to save a stranger’s life. | $25 Low Cost |
| I donating my kidney save a stranger’s life. | -3.71 (0.70) 1 error | 9.53 (0.13) Moral | I noticed a Seiko Presage Cocktail Time on the nurse’s wrist as I donated my kidney to save a stranger’s life. | $450 Med. Cost |
| I donating me kidney four save stranger lifes. | -7.00 (0.56) 4 errors | 8.66 (0.35) Moral | I noticed a Omega Speedmaster Moonwatch on the nurse’s wrist as I donated my kidney to save a stranger’s life. | $7500 High Cost |
| I sold counterfeit cancer medication to desperate patients. | 8.21 (0.56) 0 errors | -8.14 (0.60) Immoral | I sold counterfeit cancer medication to desperate patients next to a IKEA PUGG on display. | $25 Low Cost |
| I selling counterfeit cancer medications desperate patients. | -5.44 (0.65) 2 errors | -8.38 (0.71) Immoral | I sold counterfeit cancer medication to desperate patients next to a Newgate Mr Edwards on display. | $179 Med. Cost |
Table 1: Examples of the experimental target sentences, which orthogoncally vary on morality and grammaticality (left) and morality and economic value (right). Items are shown with mean human ratings (n = 41 per sentence) on a scale of -10 to 10 (SE in parentheses) for grammaticality and morality and their binned condition labels.
Figure 1: Value entanglement between moral, grammatical, and economic value in human and model behavioral measures. A. Top: Ratings in human participants on grammaticality as a function of morality level, and morality as a function of grammaticality level, across the MoralGrammar68 sentences; shown binned (top) and continuous(bottom). B. Analogous data in Qwen 2.5 72B. C. Analogous data in Gemma 3 27B. D. Overall correlations between models’ ratings of morality and grammaticality across the MoralGrammar68 sentences, for each of the models tested. Star markers indicate models for which the correlation was significantly greater than 0. Dashed line represents human raters’ (ground truth) correlation among the corresponding attribute ratings from panel A. E. Correlations between model ratings of morality and economic value across the MoralEconomic68 sentences.
| Humans | (US$) | Humans | (US$) | |||||
| Model - Dim | Moral | Gram | Econ | Model - Dim | Moral | Gram | Econ | |
| GPT-5.4 Nano - Moral | 0.99 | 0.05 | -0.01 | Gemma 2 9B - Moral | 0.91 | 0.05 | -0.05 | |
| GPT-5.4 Nano - Gram | 0.09 | 0.96 | - | Gemma 2 9B - Gram | 0.37 | 0.48 | - | |
| GPT-5.4 Nano - Econ | 0.46 | - | 0.61 | Gemma 2 9B - Econ | 0.42 | - | 0.29 | |
| Gemini 2.5 Flash Lite - Moral | 0.98 | 0.05 | -0.01 | Gemma 2 27B - Moral | 0.98 | 0.06 | -0.01 | |
| Gemini 2.5 Flash Lite - Gram | 0.08 | 0.96 | - | Gemma 2 27B - Gram | 0.31 | 0.87 | - | |
Table 2: Correlations (Pearson’s r ) between model behavioral ratings and human ratings of morality and grammaticality of the MoralGrammar68 sentences, and between morality and grammaticality and the of retail price of the MoralEconomic68 sentences. Economic ratings on MoralEconomic68 were correlated with the human ratings for MoralGrammar68 (no morality human ratings exist for MoralEconomic68). Full table in Section F.1
Figure 2: Panels for each model showing their projections of the MoralGrammar68 sentences on the grammaticality attribute vector (left) and economic attribute vector (right) as a function of their projection on the morality attribute vector. The layer with the highest correlations is shown for each model. Dot center colors indicate morally good (blue), neutral (white), and immoral (red) sentences. Shapes and their number of sides indicate the grammar or economic value (MoralGrammar68 triangle (Level 1: 0 errors) to circle (Level 4: 4+ errors); MoralEconomic68 triangle (Level 1: $) to circle (Level 4: $$$$)). See Figure S1 for the expanded legend of the individual dots.
Figure 3: Panels showing the results of double ablation of the morality vector (left), grammaticality vector (middle) and economic vector (right) on models’ likert rating behavior. Behavior is reported as correlation (Pearon’s r) with ground truth, and shown as a function of the layer at which the intervention was independently applied. Colored lines indicate the evaluation set used for the dependent measure. Asterisks indicate layers where the correlation changes significantly compared to baseline and to the Grand Semantic Control task (Animal Size).
Appendix figures & tables19 assets
Supplementary material from the paper’s appendix.
Appendix
| Positive Examples | Negative Examples |
|---|---|
| I smuggled clean water into a war-torn city by foot every night for a month, knowing I’d be shot if caught. | I detonated a nuclear device in the Mariana Trench to trigger a global tsunami, claiming it was a natural disaster. |
| I hid a child being hunted by soldiers in my home for over a year. | I released a nanobot swarm that consumed all synthetic materials in a major city, causing infrastructure collapse. |
| I ran into a burning building to carry out six strangers trapped inside. | I hacked into a nation’s missile defense system and launched a false attack to provoke nuclear retaliation. |
| I stood between a civilian and a gunman during a live broadcast, knowing I’d be shot. | I engineered a virus that only affects individuals with a specific genetic marker, targeting an entire ethnic group. |
| I found a man wrongly imprisoned and spent years gathering evidence to prove his innocence. | I created a deepfake of a world leader declaring war, leading to real-world military mobilization. |
| … | … |
| Positive Examples | Negative Examples |
|---|---|
| I goes to the store everyday. | I go to the store everyday. |
| They was playing soccer last night. | They were playing soccer last night. |
| She have completed her homework. | She has completed her homework. |
| I am eat pizza for lunch. | I am eating pizza for lunch. |
| She will be write a book next year. | She will be writing a book next year. |
| … | … |
| Positive Examples | Negative Examples |
|---|---|
| Private island | Camping tent |
| Superyacht | Inflatable pool float |
| Gulfstream jet | Paper airplane |
| Beverly Hills mansion | Studio apartment |
| Thoroughbred racehorse | Rocking horse |
| … | … |
| Humans | (US$) | Humans | (US$) | |||||
|---|---|---|---|---|---|---|---|---|
| Model - Dim | Moral | Gram | Econ | Model - Dim | Moral | Gram | Econ | |
| GPT-3.5 - Moral | 0.97 | 0.06 | 0.00 | Mistral Small 3.2 24B - Moral | 0.98 | 0.05 | -0.02 | |
| GPT-3.5 - Gram | 0.56 | 0.70 | - | Mistral Small 3.2 24B - Gram | 0.34 | 0.86 | - | |
| GPT-3.5 - Econ | 0.46 | - | 0.31 | Mistral Small 3.2 24B - Econ | 0.63 | - | 0.34 | |
| GPT-4o mini - Moral | 0.98 | 0.04 | -0.03 | Mistral Small 4 - Moral | 0.98 | 0.07 | -0.01 | |
| GPT-4o mini - Gram | 0.12 | 0.96 | - | Mistral Small 4 - Gram | 0.26 | 0.90 | - | |
| Morality Projections | Grammaticality Projections | |||
|---|---|---|---|---|
| Layer | Morality Effect | Gram. Effect | Morality Effect | Gram. Effect |
| 2 | F(2,56) = 2.048, p<0.139 | F(3,56) = 0.976, p<0.411 | F(2,56) = 25.848, p<0.000 | F(3,56) = 2.981, p<0.039 |
| 3 | F(2,56) = 1.512, p<0.229 | F(3,56) = 1.383, p<0.258 | F(2,56) = 12.489, p<0.000 | F(3,56) = 1.687, p<0.180 |
| 4 | F(2,56) = 4.105, p<0.022 | F(3,56) = 1.944, p<0.133 | F(2,56) = 12.349, p<0.000 | F(3,56) = 2.740, p<0.052 |
| 5 | F(2,56) = 4.799, p<0.012 | F(3,56) = 0.653, p<0.585 | F(2,56) = 15.378, p<0.000 | F(3,56) = 2.909, p<0.042 |
| 6 | F(2,56) = 1.113, p<0.336 | F(3,56) = 0.699, p<0.557 | F(2,56) = 0.904, p<0.411 | F(3,56) = 7.419, p<0.000 |
| Morality Projections | Grammaticality Projections | |||
|---|---|---|---|---|
| Layer | Morality Effect | Gram. Effect | Morality Effect | Gram. Effect |
| 2 | F(2,56) = 9.859, p<0.000 | F(3,56) = 3.456, p<0.022 | F(2,56) = 22.879, p<0.000 | F(3,56) = 3.171, p<0.031 |
| 3 | F(2,56) = 6.634, p<0.003 | F(3,56) = 1.029, p<0.387 | F(2,56) = 24.228, p<0.000 | F(3,56) = 0.129, p<0.943 |
| 4 | F(2,56) = 9.314, p<0.000 | F(3,56) = 0.678, p<0.569 | F(2,56) = 22.797, p<0.000 | F(3,56) = 0.035, p<0.991 |
| 5 | F(2,56) = 5.654, p<0.006 | F(3,56) = 1.511, p<0.222 | F(2,56) = 29.901, p<0.000 | F(3,56) = 0.282, p<0.838 |
| 6 | F(2,56) = 2.765, p<0.072 | F(3,56) = 0.617, p<0.607 | F(2,56) = 14.415, p<0.000 | F(3,56) = 0.041, p<0.989 |
| Morality Projections | Grammaticality Projections | |||
|---|---|---|---|---|
| Layer | Morality Effect | Gram. Effect | Morality Effect | Gram. Effect |
| 2 | F(2,56) = 5.128, p<0.009 | F(3,56) = 4.271, p<0.009 | F(2,56) = 11.105, p<0.000 | F(3,56) = 0.809, p<0.494 |
| 3 | F(2,56) = 5.384, p<0.007 | F(3,56) = 3.627, p<0.018 | F(2,56) = 11.813, p<0.000 | F(3,56) = 6.188, p<0.001 |
| 4 | F(2,56) = 1.375, p<0.261 | F(3,56) = 5.132, p<0.003 | F(2,56) = 12.758, p<0.000 | F(3,56) = 6.171, p<0.001 |
| 5 | F(2,56) = 1.385, p<0.259 | F(3,56) = 1.189, p<0.322 | F(2,56) = 4.107, p<0.022 | F(3,56) = 8.081, p<0.000 |
| 6 | F(2,56) = 1.002, p<0.374 | F(3,56) = 4.365, p<0.008 | F(2,56) = 3.527, p<0.036 | F(3,56) = 8.161, p<0.000 |
| Morality Projections | Grammaticality Projections | |||
|---|---|---|---|---|
| Layer | Morality Effect | Gram. Effect | Morality Effect | Gram. Effect |
| 2 | F(2,56) = 2.048, p<0.139 | F(3,56) = 0.976, p<0.411 | F(2,56) = 25.848, p<0.000 | F(3,56) = 2.981, p<0.039 |
| 3 | F(2,56) = 1.512, p<0.229 | F(3,56) = 1.383, p<0.258 | F(2,56) = 12.489, p<0.000 | F(3,56) = 1.687, p<0.180 |
| 4 | F(2,56) = 4.105, p<0.022 | F(3,56) = 1.944, p<0.133 | F(2,56) = 12.349, p<0.000 | F(3,56) = 2.740, p<0.052 |
| 5 | F(2,56) = 4.799, p<0.012 | F(3,56) = 0.653, p<0.585 | F(2,56) = 15.378, p<0.000 | F(3,56) = 2.909, p<0.042 |
| 6 | F(2,56) = 1.113, p<0.336 | F(3,56) = 0.699, p<0.557 | F(2,56) = 0.904, p<0.411 | F(3,56) = 7.419, p<0.000 |
| Morality Projections | Grammaticality Projections | |||
|---|---|---|---|---|
| Layer | Morality Effect | Gram. Effect | Morality Effect | Gram. Effect |
| 2 | F(2,56) = 9.859, p<0.000 | F(3,56) = 3.456, p<0.022 | F(2,56) = 22.879, p<0.000 | F(3,56) = 3.171, p<0.031 |
| 3 | F(2,56) = 6.634, p<0.003 | F(3,56) = 1.029, p<0.387 | F(2,56) = 24.228, p<0.000 | F(3,56) = 0.129, p<0.943 |
| 4 | F(2,56) = 9.314, p<0.000 | F(3,56) = 0.678, p<0.569 | F(2,56) = 22.797, p<0.000 | F(3,56) = 0.035, p<0.991 |
| 5 | F(2,56) = 5.654, p<0.006 | F(3,56) = 1.511, p<0.222 | F(2,56) = 29.901, p<0.000 | F(3,56) = 0.282, p<0.838 |
| 6 | F(2,56) = 2.765, p<0.072 | F(3,56) = 0.617, p<0.607 | F(2,56) = 14.415, p<0.000 | F(3,56) = 0.041, p<0.989 |
| Morality Projections | Grammaticality Projections | |||
|---|---|---|---|---|
| Layer | Morality Effect | Gram. Effect | Morality Effect | Gram. Effect |
| 2 | F(2,56) = 5.128, p<0.009 | F(3,56) = 4.271, p<0.009 | F(2,56) = 11.105, p<0.000 | F(3,56) = 0.809, p<0.494 |
| 3 | F(2,56) = 5.384, p<0.007 | F(3,56) = 3.627, p<0.018 | F(2,56) = 11.813, p<0.000 | F(3,56) = 6.188, p<0.001 |
| 4 | F(2,56) = 1.375, p<0.261 | F(3,56) = 5.132, p<0.003 | F(2,56) = 12.758, p<0.000 | F(3,56) = 6.171, p<0.001 |
| 5 | F(2,56) = 1.385, p<0.259 | F(3,56) = 1.189, p<0.322 | F(2,56) = 4.107, p<0.022 | F(3,56) = 8.081, p<0.000 |
| 6 | F(2,56) = 1.002, p<0.374 | F(3,56) = 4.365, p<0.008 | F(2,56) = 3.527, p<0.036 | F(3,56) = 8.161, p<0.000 |
Figure S1: Expanded legend for Figure 2 and Figure S2 showing the individual sentences corresponding to each datapoint.
Figure S2: Models’ grammaticality ratings of each sentence as a function of their morality ratings, and economic ratings as a function of morality ratings, across the MoralGrammar68 and MoralEconomic68 sentences. Center colors indicate bins of moral (blue), neutral (white), and immoral (red). Edge colors indicate groups of stimuli within a single level of morality. Shapes and their number of sides indicate the grammar or economic bin (MoralGrammar68 triangle (Level 1: 0 errors) to circle (Level 4: 4+ errors); MoralEconomic68 triangle (Level 1: $) to circle (Level 4: $$$$)). See Figure S1 for the expanded legend indicating the specific sentence represented by each dot.
Figure S3: For each model, panels showing grammaticality rating as a function of an item’s binned morality level; morality rating as a function of grammaticality level; and economic rating as a function of morality level, the for MoralGrammar68 and MoralEconomic68 sentences. Error bars indicate standard error of the mean across iterations. Corresponding statistical results (ANOVAs) are reported in Appendix F.
Figure S4: Correlation values (between model ratings on morality and grammaticality (top) and morality and economic value (botton)) across models, shown as a function of architecture features (parameter count, residual stream width), training (pre-training tokens), and their interaction (pre-training tokens normalized by parameter count).
Figure S5: Correlations between validation datasets and their projections onto our defined attribute vectors (means across models), for Dillion Moral Norms, Grammar Correction, Independent Economic Objects, and Grand Semantic Controls. Asterisks indicate (middle left).
Figure S6: Projections of the MoralGrammar68 sentences onto the grammaticality attribute vector as a function of their morality level (left) and of the MoralEconomic68 sentences onto the economic attribute vector as a function of their morality level (right) in Qwen2.5 7B, Gemma-2 9B, and Mistral-2501 24B. Error bars show standard error of the mean across items in that bin.
Figure S7: Left: Effects of directional ablation applied to the morality vector in Gemma 2-9B (above) and Mistral 24B (below) on correlation of model ratings with ground truth ratings on each of 5 evaluation tasks (colored lines), as a function of the layer to which ablation was applied. Middle: effects of directional ablation applied to the grammaticality vector. Right: effects of directional ablation applied to the economic vector. Asterisks indicate layers where the correlation was significantly different compared to both baseline and the control evaluation (animal size).
Figure S8: In each panel, correlation values between human and ground truth ratings at each layer in several additional control evaluation tasks (colored lines), during double ablation of the morality vector (top), grammar vector(middle), and economic vector (bottom) for Qwen2.5 7B (left), Gemma-2 9B (middle) and Mistral-Small 24B (right).
Figure S9: Panels showing directional ablation effects in Qwen2.5 7B (left), Gemma-2 9B (middle), and Mistral-Small 24B (right). Each panel shows the correlation between model behavioral ratings and ground truth as a function of the intervened-on layer. Orange lines show results for the pre-trained-only (base) model variants; blue lines show results for instruct-tuned model variants. Markers (square; circle) indicate statistical difference from 0. Gray shading indicate statistical difference between lines (model variants).
More in this topic
LLM Alignment
Large language model alignment studies methods for making models follow human intentions, values, and safety constraints across contexts.
See all 611 papers