Long-Horizon Robustness

Long-horizon robustness is the safety property of maintaining consistent, correct behavior — particularly around refusals and policy boundaries — across extended, multi-turn conversations. In short interactions, a model's safety behavior can be tested and verified relatively easily: each turn is independent, the context window is small, and it is straightforward to verify that the model refuses harmful requests and permits legitimate ones. In long sessions, the model's behavior can drift: context from earlier turns can bias it toward or away from certain responses, jailbreak attempts can accumulate, and the model's refusals can become inconsistent as the conversation history grows.

The challenge is particularly acute for interaction models, which are designed for exactly the kind of extended, continuous engagement where long-horizon robustness matters most. A model that is trusted to remain present in a conversation for minutes or hours — answering follow-ups, integrating background research, maintaining conversational coherence — must maintain its safety properties across that entire duration without degradation. Standard safety evaluation focuses on single-turn or short-context interactions and does not measure whether safety behavior is maintained over time.

Interaction model safety research uses an automated red-teaming harness to generate multi-turn refusal data, systematically testing whether the model's refusal behavior changes after extended conversation on sensitive topics. The harness simulates realistic multi-turn conversations that gradually escalate toward disallowed content, testing whether the model maintains its refusal boundary consistently or gradually shifts its behavior in response to accumulated context. The goal is behavioral parity with text-based safety: the model's audio refusals should be no less consistent than its text refusals across equivalent long-horizon scenarios.

Long-horizon robustness is particularly challenging for interaction models because the continuous audio and video streams create more opportunities for context accumulation than turn-based text interactions. A user's conversational framing, tone, and accumulated implicit context can influence the model's behavior in ways that are harder to detect and correct than explicit text prompt injections. Ensuring robust safety under these conditions requires both better training data (covering longer and more varied conversation arcs) and better evaluation frameworks that can detect subtle behavioral drift before it becomes a safety incident.