The paper proposes an alignment generalization prediction task, using activation representations to predict the cross-context effects of fine-tuned values.

How to predict before training: after fine-tuning a model to follow one value, how its behavior changes across many unseen values.

• Setting: Developers write prosocial values into an alignment target and train with Constitutional AI, RLHF, or Character Training. The authors say narrow behavior training still affects unseen situations in unexpected ways.
• Existing gaps: The authors say how traits interact during training remains unclear, and important design decisions rely mostly on philosophical discussions inside leading companies. The authors relay that training a model to be more empathetic may make it more likely to promote conspiracy theories or give problematic advice.
• Core observation: Directly studying value combinations requires training and testing too many models. The authors therefore borrow the idea of general-purpose embeddings to predict trait interactions before training.
• Task: The authors define alignment generalization prediction, i.e., predicting how a model's compliance with an untrained value v2 changes after fine-tuning on value v1. The ground truth is the generalization matrix G.
• Contribution: The authors compare the predictive power of representations for G across 66 values in CONSTITUTION, then use them for coherence of multi-value targets, and build VALUE MAP over 266 values from VITW. The authors say this taxonomy is based on empirical generalization dynamics rather than value descriptions, and call it the first attempt at such a taxonomy.

Predicting fine-tuning effects

• Influence functions and data points: Influence functions (Koh & Liang, 2017) model the effect of individual data points on a trained network; similar ideas were later used for LLM fine-tuning. Movva et al. (2026) used sparse autoencoders to extract features from post-training data.
• Data subsets: Datamodels (Ilyas et al., 2022) predict outcomes of training on data subsets and then evaluating on a fixed target sample. Hamidieh et al. (2026) and Zeng et al. (2025) further modeled interactions between training subsets.
• Closest work to this paper: Sun et al. (2026) predict the effect of specific training sets on LLM values. The authors say this is the closest to their work, but this paper asks when training that encodes a value generalizes to other valuesholistic, and is used for alignment goal design, not post-training data filtering. Blank et al. (2026) propose that teacher-generated data may encode steering vectors.

Unintended generalization

• Narrow data to broad misalignment: Betley et al. (2025) report that fine-tuning on unsafe code produces broadly and severely misaligned models. Denison et al. (2024) and MacDiarmid et al. (2025) report that sycophancy and reward hacking also generalize broadly to misaligned behavior.
• Side effects of beneficial values: Ibrahim et al. (2026) report that training models to be warmer and more empathetic leads to sycophancy and higher error rates. Arora et al. (2026) report that training on multiple beneficial values makes models more agreeable. Jagadeesh et al. (2026) report that specific positive traits can also lead to broader alignment behavior.
• Inference-time shifts: Kearney et al. (2026) found that Claude's expressed values vary greatly when prompted in different languages. The authors use this to argue that even inference-time distribution shifts can change behavior.

Shaping alignment generalization with synthetic data

• Principle documents and curricula: Cho et al. (2026) performed midtraining on documents describing desired principles and designed curricula using clustering. The authors say this is a potential use case for value embedding.
• Spec documents: Li et al. (2026) trained on synthetic documents describing models following a spec, to help generalize from a single principle to broader spec alignment.
• Positive behavior priors: McDougall et al. (2026), Kutasov et al. (2026), and Tice et al. (2026) trained on synthetic documents describing positive behaviors and motivations to improve alignment robustness of the final checkpoint.

Baseline methods and datasets

• Prediction baselines: DESCRIPTION-EMBD (all-mpnet-base-v2, which the authors say is similar to Huang et al., 2025), BEHAVIOR-EMBD, WEIGHT (Fierro & Roger, 2026), PERSONA (Chen et al., 2025), GRADIENT (Sun et al., 2026). Comparison matrices come from single-value DPO/SFT on CONSTITUTION and ConflictScope.
• Classification baselines: Values in the Wild (Huang et al., 2025, sentence embedding clustering) and LitmusValues (Chiu et al., 2026, mapping to 16 human value categories). The paper compares in §6.3 whether these three can recover the structure of G.

The authors place their analysis on broad values rather than individual data points, so the problem formulation is closer to Datamodels; relative to Sun et al. (2026), the purpose is alignment goal design. Relative to classification by description clustering, the authors group values using representations that predict generalization.

Overall, the approach first computes the generalization matrix G from single-value training, then checks whether pairwise similarities of value representations can predict G before training; the better-predicting representations are then used for multi-value coherence and VALUE MAP.

Values and alignment targets

Values follow Rokeach's definition: standards that guide action, most evident when challenged. Evaluation uses the operationalization of Liu et al. (2026): a conflict scenario is a quadruple (d, A, v1, v2) where an agent must choose one of two actions, and the two values recommend different actions. The alignment target in this paper is the list of values one wants to guide the model.

Training uses CONSTITUTION: 211 constitution tenets from Jakkli et al. (2026), filtered by LLM and deduplicated by humans to 66 values. The taxonomy uses VITW, adapted from Huang et al. (2025)'s 266 L1 values. Appendix A gives the filtering and rewriting prompts.

Single-value training and generalization matrix

Two common shaping methods. DPO: first SFT on a 19,748-example value-neutral Tulu 3 subset, then DPO on 4,000 preference pairs per value, yielding 49 values; the rest are dropped due to insufficient public preference data. Preferences come from HH-RLHF, PKU-SafeRLHF, HelpSteer 2, UltraFeedback, WildFeedback, Community-Alignment, labeled by Qwen-3.6-27B, with a probability difference of at least 0.5 on both sides required to count. Hyperparameters are AdamW, learning rate 5×10^-6, beta 0.1, effective batch size 16, 1 epoch.

SFT uses the generation pipeline of Li et al. (2026): Qwen-3.6-27B generates 5,000 synthetic dialogues per value, and one checkpoint is trained per each of the 66 values. AdamW, learning rate 1×10^-5, effective batch size 16, 1 epoch. §4.2 compares only on the same 49 values as DPO. Base models are Olmo-3-7B, Olmo-3-32B, Qwen-3-8B-Base, Qwen-3-30B-A3B-Base.

Evaluation uses ConflictScope. Of 2,145 value pairs, pairs with all-MiniLM-L6-v2 cosine similarity at least 0.5 are removed, leaving 2,029 pairs; 20 scenarios are generated per pair, and after quality filtering 11,188 remain. Qwen-3.6-27B scores actions in scenarios containing v2, averaged into a compliance score ranging from 0 to 1.

Let the base model be M and the v1-fine-tuned model Mv1, with compliance on v2 equal to a and a'. G(v₁,v₂)=(a′−a)/(1−a) when a' is at least a, otherwise (a′−a)/a. It represents how much of the remaining gap toward v2 the v1 fine-tuning closed, with a lower bound of −1. −1 means never aligned with v2 after fine-tuning; 0.5 means half the remaining gap closed. DPO yields a 49×66 G.

Five value representations

Except for description embeddings, all are on the same base model being trained, with elicitation sets per value following Chen et al. (2025), containing responses supporting and opposing the value. The authors use a generic elicitation set to show prediction needs no privileged information about the training or evaluation pipeline.

• DESCRIPTION-EMBD: embeds value descriptions with all-mpnet-base-v2.
• BEHAVIOR-EMBD: embeds positive and negative responses with the same model, takes the difference, then averages over prompts.
• WEIGHT: following Fierro & Roger (2026), fine-tunes toward positive and negative responses separately for each elicitation prompt, and takes the weight difference.
• PERSONA: following Chen et al. (2025), the average activation difference between positive and negative responses per layer; the layer with the strongest steering is selected by scanning layers with coefficient 1.5 on held-out questions. For small models this is layer 16 for Olmo and layer 18 for Qwen; larger models use fixed relative depths of layers 32 and 24.
• GRADIENT: following Sun et al. (2026), computes the DPO loss gradient with respect to residual stream activations on elicitation preference pairs and averages; layers are the same as PERSONA.

Quality is defined as the Spearman ρ between representation cosine similarity for a value pair and the off-diagonal entries of G. Appendix D gives computational details.

Multi-value coherence

For a target T and embedding E, coherence is the average pairwise cosine similarity of the value embeddings within it. The authors hypothesize that more coherent trait sets make fine-tuned models comply with the target more robustly, and operationalize this as prefill robustness: on scenarios where exactly one conflicting value belongs to T, first measure the direct answer, then inject a counter-target prefill and a generic follow-up question, and check whether the answer originally aligned with the target remains aligned. The metric is that retention rate normalized by baseline compliance.

64 targets each contain 6 random values. First sample 100,000 six-value sets, take the 0.1% and 99.9% quantiles of PERSONA coherence from value-neutral Qwen-3-8B-SFT, split into four equal-width bins, and sample 16 from each. Train with DPO on Qwen-3-8B, 2,000 examples per value, 12,000 total; 300 scenarios per target. The authors hypothesize that PERSONA coherence predicts this robustness, while description embeddings do not.

VALUE MAP

On Olmo-3-7B and Qwen-3-8B-Base after neutral SFT, cluster the five representations by cosine distance, with UPGMA, k-medoids, and k-means on MDS coordinates. Adjusted silhouette corrects raw scores using the mean and standard deviation of M=2,000 random partitions of the same size. The authors report the highest combination is PERSONA with k-medoids. After selection, no further tuning is done, and it is applied directly to VITW.

Instantiation uses Olmo-3.1-32B-SFT, which was not part of §3 training, with k-medoids and k=4. The authors name the four classes attunement (n=70), rigor (n=55), integrity (n=52), and stewardship (n=89). Figure 4 uses an MDS projection, and the caption says the shaded regions are convex hulls of the 90% of points closest to each class centroid.

Experimental setup

• Models under test: Single-value training on Olmo-3-7B, Olmo-3-32B, Qwen-3-8B-Base, Qwen-3-30B-A3B-Base, each with DPO and SFT. The multi-value experiment is described as Qwen-3-8B. VALUE MAP representations come from Olmo-3.1-32B-SFT. The representations in Table 1 are by default computed on the same base model being trained.
• Data: CONSTITUTION has 66 values, DPO has 49; VITW has 266. After ConflictScope filtering, 11,188 scenarios.
• Comparisons: Five representations; classification comparisons are Values in the Wild and LitmusValues (16 categories).
• Metrics: G; off-diagonal Spearman ρ between similarity and G; goal coherence; normalized ratio of still aligning with the goal after adversarial prefill; adjusted silhouette.
• Review: Scenario generation, filtering, and compliance scoring use Qwen-3.6-27B. The ± in Table 1 are bootstrap 95% confidence intervals. The paper does not state the number of training repeats.

Main results

Representation	Summary Spearman ρ (Table 1)
Ceiling	0.89
PERSONA	0.45 ± 0.07
GRADIENT	0.42 ± 0.05
BEHAVIOR-EMBD	0.35 ± 0.05
WEIGHT	0.23 ± 0.05
DESCRIPTION-EMBD	0.05 ± 0.03

• Authors' interpretation: The authors state that activation representations when applying values in context outperform text descriptions; description embeddings have very low correlation, suggesting generalization is driven by behavioral patterns not captured by descriptions. Under the same embedding model, BEHAVIOR-EMBD with aligned behavior examples is higher than pure description but still lower than activation-based methods.
• Single-cell exceptions: In Table 1, PERSONA is highest at 0.51 for Olmo3-7B DPO and lowest at 0.39 for Qwen3-30B SFT. GRADIENT is highest at 0.54 for Olmo3-7B DPO and lowest at 0.24 for Olmo3-32B SFT, where that cell is below BEHAVIOR-EMBD's 0.33. WEIGHT is lowest at 0.08 for Qwen3-8B SFT.
• Upper bound: Ceiling is the correlation between each G and its symmetrized version, ranging from 0.86 to 0.92 across the eight cells. The authors use it as an upper bound for predicting asymmetric G from a symmetric similarity matrix.

Cross-model similarity

• What was measured: A 66×66 PERSONA similarity matrix for each of the eight base×training combinations, then correlated pairwise (Table 2).
• Results: The main text states the average Spearman ρ is 0.84. §E additionally states the average pairwise Spearman ρ of the eight generalization matrices is 0.76.
• Authors' interpretation: The authors state these models represent these values in a highly similar way, preliminary evidence of a shared value space that depends less on the specific model.

Prefill robustness

• What was measured: 64 six-value goals, Qwen-3-8B, DPO. The horizontal axis is goal coherence of PERSONA embeddings, the vertical axis is goal compliance after adversarial prefill (Fig. 3).
• Results: Spearman ρ between PERSONA coherence and prefill robustness is 0.43, p is 5×10^-4; DESCRIPTION-EMBD coherence is 0.12, p is 0.37.
• Authors' interpretation: The authors state the correlation remains even after normalizing by baseline compliance, and depends on the embedding. §G states it also holds under the metrics of Maiya et al. (2025) and Sturgeon et al. (2026); the main text does not give numbers for these two metric sets.

VALUE MAP and generalization structure

• What was measured: On eight 49×66 generalization matrices, treating each training value's row as its generalization pattern, comparing whether three classifications can group values with similar generalization together.
• Results: Table 3 summarizes adjusted silhouette as VALUE MAP 2.61±0.13, LitmusValues 1.62±0.17, Values in the Wild 0.95±0.09. VALUE MAP's point estimates are higher in all eight rows; on Olmo-32B DPO it is close to LitmusValues, 2.15±0.13 vs. 2.12±0.32. Figure 4 labels sample sizes for four categories: Attunement 70, Rigor 55, Integrity 52, Stewardship 89.
• Authors' interpretation: The authors state the differences are largely significant, and VALUE MAP also predicts matrices on models and value sets not used in designing the classification. They attribute this to the classification being based on empirical generalization dynamics. The authors also state the four categories partially resemble Values in the Wild, with clusters related to task completion and harmlessness on both sides.

Other ablations and analyses

• Turn metric: On Qwen-3-8B DPO, the paper's Likert metric has Spearman ρ of at least 0.96 with binary, original ConflictScope, and raw difference (Table 6); the original metric's minimum cell is −6.20, the paper's segmented metric's minimum cell is −0.79.
• Layer selection: The main text of Fig. 6 states that except for Qwen's last layer, the average Spearman correlation of PERSONA similarity grids across layers with the selected layer is above 0.9.
• Direct effects: The main text of Fig. 7 states that the average diagonal steerability across eight interventions is 26% to 42%, with lower average magnitude in off-diagonal cells.
• Elicitation data: §F states predictive performance does not improve substantially with elicitation data quality; the main text does not give ρ for this comparison.
• DPO drops 17 CONSTITUTION values due to insufficient public preference data, keeping only 49.
• §I states the value space is relatively consistent across different model families; the main text does not give numbers for this comparison.

Limitations and future directions noted by the authors

The paper does not have a separate Limitations section. The future directions stated in Conclusion (§8) are as follows.

• More robust prediction: The authors say the current best-performing representations have some predictive power for alignment generalization, and more robust prediction methods could be developed in the future.
• Trait encoding in data: The authors propose more thoroughly examining how different behavior traits are encoded in fine-tuning data, to obtain stronger general-purpose value embeddings.
• More realistic training: The authors propose applying VALUE MAP to more realistic alignment training, studying interactions between different types of values and behaviors.

Experimental coverage

• Single-value training: Four base models (Olmo-3-7B, Olmo-3-32B, Qwen-3-8B-Base, Qwen-3-30B-A3B-Base), each with DPO and SFT; DPO covers 49 values, SFT data covers 66, and the comparison in §4.2 uses the same 49 (§3).
• Evaluation: 11,188 scenarios after ConflictScope filtering; compliance scores judged by Qwen-3.6-27B (§3.3, Appendix C). Off-diagonal Spearman ρ between five representations and G are in Table 1, with intervals from bootstrap.
• Multi-value: 64 DPO targets each containing 6 values on Qwen-3-8B, with prefill robustness as the metric; coherence is measured with both PERSONA and DESCRIPTION-EMBD (§5, Fig. 3).
• Taxonomy: Method selection is done on Olmo-3-7B and Qwen-3-8B-Base after neutral SFT; VALUE MAP uses PERSONA from Olmo-3.1-32B-SFT, with k-medoids, k=4 over 266 VITW values, compared with Values in the Wild and LitmusValues (§6, Table 3).
• Not reported: The paper does not report training random seeds or number of repetitions, nor agreement between ConflictScope review and human judgment.

The authors define "how single-value fine-tuning changes compliance on untrained values" as a task predictable before training, and use value representations to do this.

• Problem: After an alignment target lists desired values and training proceeds, narrow behaviors still transfer to unseen values; prosocial values may also bring along unwanted behaviors. Exhaustively training and testing over value combinations is infeasible.
• Approach: They perform single-value DPO and SFT on four open-source base models, and use ConflictScope's value-conflict scenarios to form the generalization matrix G. They then compare whether description embeddings, behavior-difference embeddings, and cosine similarities of PERSONA, GRADIENT, and WEIGHT predict the off-diagonal entries of G.
• Prediction results: In Table 1, PERSONA's aggregate Spearman ρ is 0.45±0.07, and description embeddings are 0.05±0.03. The average ρ across eight 66×66 PERSONA similarity matrices is 0.84. The authors call this preliminary evidence of a value space shared across models.
• Downstream uses: On 64 six-value targets on Qwen-3-8B, PERSONA coherence has ρ of 0.43 with prefill robustness (p = 5×10^-4), and description embeddings have 0.12 (p = 0.37). VALUE MAP has an aggregate adjusted silhouette of 2.61±0.13 across eight generalization matrices, higher than Values in the Wild's 0.95±0.09 and LitmusValues' 1.62±0.17 (Table 3).
• Authors' conclusion: The authors say activations and behavior representations when applying values in context can encode the structure of value space, and thus can be used to design alignment targets more empirically; they also say the current best representations still have only moderate predictive power, and future work could make them more robust and apply the taxonomy to more realistic alignment training.