arXiv:2510. 26707v2 Announce Type: replace-cross Abstract: As LLMs occupy an increasingly important role in society, they are more and more confronted with questions that require them not only to draw on their general knowledge but also to align with certain human value systems.
The paper investigates the mismatch between large language models’ verbalized (knowledge) distributions and their actual output (behavior) distributions when predicting group opinions. It shows that for 37% of World Values Survey questions, the most likely answer under the verbalized distribution differs from that under the log‑probability distribution, indicating a disconnect between what the model knows and what it outputs. To reconcile this, the authors propose a learnable combination of the two distributions that preserves the knowledge shape while aligning with the behavior’s most likely answer, achieving higher behavioral consistency with only a modest increase in distance to human opinion distributions.
arXiv:2608.30902v1 Announce Type: new Abstract: Adapting large language models to user-specific preferences is often constrained by the cost of human annotation, making preference optimisation imprac...
arXiv:2608.29803v1 Announce Type: cross Abstract: Large language models (LLMs) are increasingly deployed as proxies for human participants in social simulations, yet whether they update their beliefs...
The paper investigates how large language models develop brand preferences, noting that models often name market‑leading brands like Nike when asked for a running shoe brand but recommend niche brands such as Brooks. Across 102 models and 42 brand categories, the brand chosen for recommendation differs from the named brand in 26% of cases, and preference optimization (DPO) shifts these choices by 11–24 points. The study suggests that pretraining already links recommended brands to advice, supervised fine‑tuning preserves these preferences, and DPO can shift them even when the preferred brand is absent from the preference data, indicating that the judge models used in training play a key role.
Asked to name a running shoe brand, language models say Nike. Asked to recommend one, they say Brooks. This paper asks where models learn that taste. Across 102 models and 42 brand categories, the bra...
arXiv:2609.38792v1 Announce Type: new Abstract: We study training LLM judges from natural language feedback, especially for subjective tasks where the verdict depends strongly on which evaluation cri...
The paper introduces Style‑Debiased DPO (SD‑DPO), a method that refines large language models’ ability to retrieve stored knowledge by using preference optimization that corrects for style differences while preserving factual accuracy. SD‑DPO evaluates on the EntiGraph storing‑side framework and outperforms baseline CPT on the QuALITY reading‑comprehension benchmark, achieving higher accuracy with far fewer training tokens. In a knowledge‑editing setting (AToKE), SD‑DPO attains an overall accuracy of 0.982, correctly answering queries with either new or old facts based on the requested time period.
The paper investigates whether large language models (LLMs) possess intrinsic value systems and how to quantify and align them. By projecting responses from 106 LLMs and 95,000 human survey profiles into a shared sociological space, the authors confirm that LLMs do have values, though these values form a concentrated, idealized core rather than mirroring human diversity. They introduce the Prior-Environment-Cognition (PEC) framework to mathematically define value expression and propose an adaptive Alignment Prescription that identifies minimal interventions—ranging from prompts to targeted parameter updates—to steer LLM values efficiently without harming general performance.
The paper investigates whether fine‑tuning large language models (LLMs) on value survey responses can shift their value‑conditioned outputs and downstream behavior. Researchers first estimate baseline value profiles for several open‑source LLMs using the ten Schwartz values, then fine‑tune the models on survey‑style supervision that selectively alters target value ratings while preserving non‑target values. The resulting shifts are evaluated on held‑out survey questions and on out‑of‑domain behavioral proxies, such as a contextual moral‑judgment benchmark from Reddit AITA posts and a text‑based interactive decision‑making benchmark. The study finds that survey‑based supervision can change both survey responses and downstream behavior, but the transfer is strongly model‑ and value‑dependent and influenced by how value descriptions are framed.
arXiv:2609. 11291v1 Announce Type: new Abstract: We post-train Qwen3.
The paper examines how fine‑tuning large language models (LLMs) for specific human values can unintentionally affect other, untrained values and broader behaviors. By adapting four instruction‑tuned LLMs to Hofstede’s value dimensions, the authors evaluate cross‑value transfer and observe side effects on risk preferences, social cognition, and safety traits. They conclude that post‑training value alignment should include comprehensive behavioral assessments and that further research is needed to determine the optimal placement of value training within the development pipeline.