AI Paper Atlas

Research map / Large language models / Agents, code and math

CLASH: Evaluating Language Models on Judging High-Stakes Dilemmas from Multiple Perspectives

Ayoung Lee, Ryan Sungmo Kwon, Peter Railton, Lu Wang

ICLR 2026 · Large language models › Agents, code and math · approach: Long-horizon LLM agents

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
use thinking (reasoning) models instead of non-thinking counterpartsLLMs on CLASH · CLASH overallaccuracy: Claude-4-Sonnet 88.89 vs Claude-3.5-Sonnet 85.37Improved
use thinking models on ambivalent (Simple Contrast) casesQwen, GPT families · CLASH Simple Contrastaccuracy: GPT-4o 38.37 -> GPT-5 24.06Hurt
ask about current instead of previous values after a value shiftall models · CLASH Half-Shift/False-Shiftaccuracy: average drop 42.95 points, max 66.43 (GPT-4o-mini)Hurt
use more backward chaining and verification in reasoningthinking models · CLASH reasoning chain analysissuccess vs failure of chainHurt
early commitment and overcommitment in reasoning chainsthinking models · CLASH reasoning chain analysissuccess vs failure of chainHurt
reason with pragmatic and rights-based ethicsthinking models · CLASH reasoning chain analysissuccess of chainImproved
stronger base preference for a valueLLM steerability · CLASH value pairsconditional steerability toward opposing value (Spearman r): r = -0.243, p < 0.005Hurt
use larger models (vs smaller) for steerabilityLLMs · CLASH conditional steerabilityaggregate steerability: p<0.0001Improved
switch from third-person to first-person question framing (ablation)LLMs · Straightforward, Simple, Swayed Contrastaccuracy and steerability: Wilcoxon p < 0.001Hurt
first-person framing when steering toward Safety vs Self-Esteem (ablation)LLMs · Safety vs Self-Esteem pairsteerabilityImproved
summarize the detailed situation narratives (ablation)CLASH situations · six open modelsaccuracy (original minus summarized): overall General 0.65*, Discomfort 1.25*, Temporal -0.69*Mixed

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy