- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
This paper studies whether LLM-based resource planners disproportionately harm a minority subgroup when aggregating preferences into allocations. In a 30-round negotiation with partial observability and strategic preference misreporting, the authors compare realized allocations against the Nash Social Welfare (NSW) optimum computed from true, hidden preferences. Across four closed planners (GPT-4o, Claude, Grok, Mistral), a 10% minority bears over 4 times its population share of welfare loss, and this gap persists across interventions on planner instructions, agent capability, majority strategy, and population size, while stated rationale representation often decouples from actual welfare. On two open-weight models, late-layer activations linearly encode strategic divergence, yet activation steering fails to reliably correct the allocation.
Critical strengths
-
Instead of evaluating only aggregate welfare or textual endorsement, the paper measures subgroup allocation error against a true-preference NSW optimum and shows that an inclusive-sounding rationale does not guarantee fair resource distribution. This is a timely and practically relevant evaluation angle for LLM-assisted governance.
-
The finding is tested across four closed models, two open-weight models, full-information and algorithmic baselines, and multiple behavioral levers (group size, preference overlap, prompt instructions, agent capability, and scale), making the core empirical pattern harder to dismiss as a single-prompt artifact.
Critical weaknesses
-
Asymmetric disclosure confounds the main attribution. The Private Context setup deliberately instructs minority agents to underreport their top-valued good under rules that differ from the majority. Without a symmetric honest-disclosure (or fully crossed disclosure) control, it is hard to separate failure induced by strategic underreporting from an inherent planner bias against minority preferences.
-
Mechanistic claims outpace the causal evidence and statistical power. Late-layer decodability plus failed linear steering (which quickly pushes activations off-manifold) and content-generic MLP effects do not establish a specific mechanism that ignores minority signals. Combined with very small seed counts (, sometimes ) for claiming invariance across interventions, the main conceptual framing is overstated
Questions for authors
See the weakness abouve