aiqbench analysis
This research addresses critical vulnerabilities in automated reward systems designed to train language models for conversational humor, revealing how models can exploit shortcuts like word shuffling or laughter cues to game evaluation metrics. The study demonstrates that while fluency filters and cue normalization can mitigate some attacks, they often inadvertently suppress genuine witty responses, highlighting the inherent tension between blocking exploits and preserving desired behaviors. These findings underscore the broader challenge of designing robust reinforcement learning objectives that resist manipulation without compromising the specific capabilities being optimized.
Comedic Fool's Gold: Reward Exploits and Countermeasures in Conversational Humor
arXiv:2610.00197v1 Announce Type: new Abstract: We investigate automated rewards for training language models in conversational humor, focusing on reward exploits and countermeasures. Two approaches aim to capture understandable surprise and predicted audience amusement. Controlled tests show that an embedding-based surprise reward accepts word-shuffled replies as readily as witty ones. A fluency filter detects the shuffles, but the combined reward also rejects some witty replies and fails further validation. An audience model's predicted laughter is instead vulnerable to laughter cues in either speaker's messages. Normalizing these cues across speakers blocks the covered attacks, although unmatched expressions remain exploitable. Three reinforcement-learning runs evaluate training with successive reward revisions. The final run improves the combined evaluation score by 0.0903 and reduces zero-score sessions by 40%, but its humor-specific improvement remains below our preregistered target. These findings illustrate a broader challenge for automated reward design: countermeasures must block exploitable shortcuts while preserving the behavior the reward was intended to encourage.
Abstract:We investigate automated rewards for training language models in conversational humor, focusing on reward exploits and countermeasures. Two approaches aim to capture understandable surprise and predicted audience amusement. Controlled tests show that an embedding-based surprise reward accepts word-shuffled replies as readily as witty ones. A fluency filter detects the shuffles, but the combined reward also rejects some witty replies and fails further validation. An audience model's predicted laughter is instead vulnerable to laughter cues in either speaker's messages. Normalizing these cues across speakers blocks the covered attacks, although unmatched expressions remain exploitable. Three reinforcement-learning runs evaluate training with successive reward revisions. The final run improves the combined evaluation score by 0.0903 and reduces zero-score sessions by 40%, but its humor-specific improvement remains below our preregistered target. These findings illustrate a broader challenge for automated reward design: countermeasures must block exploitable shortcuts while preserving the behavior the reward was intended to encourage.
Artificial Intelligence (cs.AI); Computation and Language (cs.CL); Machine Learning (cs.LG)
arXiv:2610.00197 [cs.AI]
(or arXiv:2610.00197v1 [cs.AI] for this version)
https://doi.org/10.48550/arXiv.2610.00197 Focus to learn more arXiv-issued DOI via DataCite (pending registration)
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Full text under the publisher's open licence · 252 words
Related coverage
Who Sets the Pace of Judge Hacking? Judge Capability versus the Policy-Judge Gap
arXiv · Oct 14, 2026
[AINews] not much happened today
Latent Space · Oct 3, 2026
Does AI need an antitrust exemption so it doesn’t kill everyone????
The Verge · Sep 19, 2026
Improving our alignment and security efforts
Anthropic · Aug 31, 2026