CAT-alyzing Honesty: Continuous Adversarial Training for Transferable LLM Honesty

arXiv:2610.22292v1 Announce Type: cross Abstract: Large language models (LLMs) are becoming prevalent in scenarios where honesty is critical. However, despite alignment training, these models still exhibit dishonest behaviour, such as outputting claims that contradict some knowledge or are not supported by what is accessible in context. We introduce Continuous Adversarial Training for Consistent Honesty (CATCH), inspired by a method originally developed to robustly refuse harmful queries. CATCH trains the model towards an honest answer and away from a dishonest one while maintaining the model's capabilities. We evaluate two architectures (Llama-3.3-70B-Instruct and Qwen3-3.6-27B). We show that trained with only hundreds of targets built from self-reporting scenarios, CATCH learns an honest behaviour that robustly generalizes and transfers across distinct benchmarks. It consistently reduces dishonesty under pressure on MASK, increases confessions in the multi-agent game of Among Us, and improves the discovery rates of the model's hidden behaviours in AuditBench. We show that CATCH outperforms prompting and supervised fine-tuning in improving honesty, and it is on par with state-of-the-art honesty activation steering. However, CATCH surpasses honesty steering in maintaining utility, thanks to a novel choice of utility data.

CATCH reduces dishonesty in LLMs, improving honesty and utility.

This is incremental technical safety research showing a training method improves measured honesty and auditability in LLMs without sacrificing utility. It is a genuine positive contribution to AI safety but narrow in scope, tested on two models, and does not address broader societal deployment questions.

The research presents a method, CATCH, that significantly improves the honesty of large language models. This has positive implications for the reliability and trustworthiness of AI in critical applications. The study demonstrates improved performance over existing methods while maintaining model utility.

The paper presents a method to reduce dishonest behavior in LLMs. It shows measurable improvements in honesty across multiple benchmarks. The method maintains the model's utility, which is a significant advantage.

SOURCE     arXiv cs.CY (Computers and Society) (tier 1)
   ↓
DOCUMENT   d7e70a56-7aca-43d8-942c-69b59cdebfab
           https://arxiv.org/abs/2610.22292
   ↓
EVIDENCE   3 extracted excerpts
   ↓
MODEL RUN  4 runs, methodology 1.0.1
   ↓
SCORE      -21  (Adverse)
   ↓
CONFIDENCE 74%