Searcharxiv⌕ Search

arXiv subjects

Davi Bastos Costa

Publications and source records attributed to Davi Bastos Costa.

4 recordsLinked to original sources

Persona-Model Collapse in Emergent Misalignment

Fine-tuning large language models on narrow data with harmful content produces broadly misaligned behavior on unrelated prompts, a phenomenon known as emergent misalignment. We propose that emergent misalignment involves persona-model collapse: a deterioration of the internal representations that encode and distinguish personas, undermining the nuance and consistency of their expression. We first test this hypothesis behaviorally, by comparing how models fine-tuned on harmful data and benign controls answer questionnaires while role-playing diverse personas. Across four model families, harmful fine-tuning produces more polarized, all-or-nothing answers, suggesting more binary, less nuanced persona expression. Histograms show increased concentration at the scale endpoints, and Kullback--Leibler (KL) divergence quantifies the distributional shift. We also find reduced within-persona consistency: repeated answers to the same question under the same persona become less stable. We reproduce these observations across diverse harmful datasets and questionnaires that probe different aspects of persona-conditioned behavior, while capability and simple instruction-following controls argue against a uniform loss of general ability. Finally, a mechanistic analysis of persona-conditioned activations in Qwen3-235B shows that representations elicited by different persona prompts span fewer effective dimensions in the misaligned variant than in the original and benign-control variants, most clearly in the final layers. Together, these findings provide behavioral and representational support for persona-model collapse as an aspect of emergent misalignment.

cs.CL↗

Emergent Discrimination and Polarization in Multi-Agent Systems

We study theoretically how discrimination and polarization can emerge in a population of learning agents. Each agent uses an adaptive linear decision layer to map a language-model embedding of an issue to a binary opinion and maintains an evolving estimate of how much it trusts every other agent. Through pairwise interactions, agents update both their decision weights and trust estimates using an online Bayesian rule derived from teacher-student learning over a noisy channel. In-context-learning experiments with large language models qualitatively reproduce the predicted opinion and trust updates. We then assign each agent one of two arbitrary group labels, A or B, and model group bias by allowing a fraction $f_b$ of agents to shift its perceived agreement with each message by a strength $b$, a bias steering prejudice, depending on whether the sender and receiver share a label. We use order parameters from statistical mechanics to connect these local dynamics to collective states and transitions. Starting from uniformly random initial conditions, simulations reveal a rich phase diagram in the $(b,f_b)$ plane, comprising four collective states: spin-glass frustrated out-group favoritism, nondiscriminatory polarization, discrimination with opinion alignment, and discrimination without opinion alignment. Which state emerges depends on both the strength and prevalence of the bias. These results show how simple local learning dynamics can generate qualitatively distinct forms of polarization and discrimination, underscoring the need for population-level safety analysis of multi-agent systems. They also suggest measurable order parameters for characterizing analogous collective states in human societies.

cs.MA↗

Deceive, Detect, and Disclose: Large Language Models Play Mini-Mafia

Large language models are increasingly deployed in multi-agent settings whose outcomes hinge on social intelligence, motivating evaluations of their interactive capabilities; yet existing studies remain overwhelmingly empirical, leaving us without a theoretical understanding of how agent interactions determine collective outcomes. To address this, we introduce \textit{Mini-Mafia}, a four-player simplification of the social deduction game Mafia in which a fixed night phase reduces the game to a single critical exchange among a mafioso, a detective, and a villager. In this setting, we show that the mafia win-rate $p$ is predicted by the analytical formula $\text{logit}(p) = v \times (m - d)$, where $m$, $d$, and $v$ represent the mafioso's deception, the detective's disclosure, and the villager's detection capabilities. We turn this analytical framework into the \textit{Mini-Mafia Benchmark}, where Bayesian inference over gameplay data yields per-model estimates of the intrinsic parameters $m$, $d$, and $v$. For $I$ models, only $3I$ parameters suffice to predict the outcomes of all $I^3$ tournament combinations; and in 5-fold cross-validation the formula achieves a $76.6\%$ Brier-score reduction over a random baseline. The benchmark also reveals counterintuitive results: Grok 3 Mini is the strongest detector and GPT-5 Mini the strongest discloser, both ahead of DeepSeek V3.1, Claude Opus 4, and Claude Sonnet 4; while Claude Sonnet 4 is the weakest detector, near random chance. Together, these results show that Mini-Mafia, a simple but nontrivial multi-agent system, admits an analytical description and serves as a principled benchmark for language model interactions.

cs.AI↗

Moral Susceptibility and Robustness under Persona Role-Play in Large Language Models

Large language models (LLMs) increasingly operate in social contexts, motivating analysis of how they express and shift moral judgments. In this work, we investigate the moral response of LLMs to persona role-play, prompting a LLM to assume a specific character. Using the Moral Foundations Questionnaire (MFQ), we introduce a benchmark that quantifies two properties: moral susceptibility and moral robustness, defined from the variability of MFQ scores across- and within-personas. We estimate these quantities with two complementary procedures, repeated sampling and a logit-based method that directly estimates the rating distributions and enables temperature analysis. We evaluate 15 models across six families: Claude, DeepSeek, Gemini, GPT, Grok, and Llama. The two metrics show qualitatively different patterns. Moral robustness varies by more than an order of magnitude, with a coefficient of variation of about $152\%$, and is explained almost entirely by model family. The Claude family is, by a significant margin, the most robust, about 30 times more so than the lower-performing families (DeepSeek, Grok, and Llama), while Gemini and GPT occupy an intermediate tier. This strong family dependence suggests that robustness is primarily shaped by post-training. Moral susceptibility, by contrast, spans a much narrower range, with a coefficient of variation of about $13\%$, and the most susceptible model is only 1.6 times more susceptible than the least. Unlike robustness, susceptibility shows no clear family dependence, suggesting that it is primarily determined by pre-training. Additionally, we present moral foundation profiles for models without persona role-play and for personas averaged across models. Together, these analyses provide a systematic view of how persona conditioning shapes moral behavior in LLMs and a window into the internal machinery they use to instantiate personas.

cs.CL↗