SPAR Fellowship unveils MLPCT and AttCT for AI consistency training

Consistency training ensures AI models maintain aligned behavior even when prompts are adversarially wrapped. Prior work introduced BCT (output-level consistency) and ACT (residual-stream consistency) to tackle sycophancy and jailbreaks. This new study, led by Sukrati Gautam, Neil Shah, Arav Dhoot, and colleagues at the SPAR Fellowship and accepted at AI4GOOD @ ICML 2026, broadens the scope. They propose two novel methods: MLPCT, which enforces consistency on post-nonlinearity MLP hidden states via cosine distance, and AttCT, which targets per-head attention distributions. Both are trained by running paired forward passes on clean and wrapped prompts, then minimizing disagreement at these internal components.

Testing against six threat models—including four new ones (persona in-context learning, prefill attacks, frustration, conditional misalignment)—the researchers uncover critical trade-offs. BCT effectively defends against prefill and persona ICL attacks but is less suited for others. Representation-level methods (MLPCT, AttCT, ACT) tend to converge on similar residual-stream representations, yet they can either entirely degrade or suppress benign behavior alongside the threat. In contrast, BCT shows distinct behavior: it reduces expressions of frustration in Gemma models and mitigates leaky conditional misalignment without sacrificing utility. The takeaway: no one-size-fits-all solution exists; the choice of consistency target and threat model dramatically impacts effectiveness.