AI
Comedic Fool's Gold: Reward Exploits and Countermeasures in Conversational Humor
Publication: cs.AI updates on arXiv.org
Author: Sam Larson
Source published: 02/10/2026, 04:00:00 UTC
Briefing created: 02/10/2026, 04:21:22 UTC
Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 02/10/2026, 04:03:24 UTC.
Briefings
This article investigates automated rewards for training language models in conversational humor, specifically focusing on reward exploits and countermeasures. Controlled tests reveal that embedding-based surprise rewards accept word-shuffled replies as readily as witty ones, while an audience model's predicted laughter is vulnerable to laughter cues in either speaker's messages. Normalizing these cues across speakers blocks covered attacks, though unmatched expressions remain exploitable. Three reinforcement-learning runs evaluated training with successive reward revisions, with the final run improving the combined evaluation score by 0.0903 and reducing zero-score sessions by 40%, though humor-specific improvement remains below the preregistered target.
Key points
- Embedding-based surprise rewards accept word-shuffled replies as readily as witty ones, while an audience model's predicted laughter is vulnerable to laughter cues in either speaker's messages.
- Normalizing laughter cues across speakers blocks covered attacks, though unmatched expressions remain exploitable.
Context and limitations
Coverage is limited to three reinforcement-learning runs evaluating training with successive reward revisions, and the final run's humor-specific improvement remains below the preregistered target.
Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 02/10/2026, 04:21:22 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.
View archived briefing on IPFS
bafybeibux4dyoj46iqcb4npunxgyhumz5md55nbrbyrwoyml6lswsjlyaq