Jane Street

Featured work

Surfacing Pathological Behaviors in Language Models

Many language model behaviors occur rarely and are difficult to detect in pre-deployment evaluations. As part of our ongoing efforts to surface and measure them, we propose a way to lower bound how often and how much a model's responses satisfy specific criteria expressed in natural language — which we call the PRopensity BOund (PRBO).

We train reinforcement learning agents to craft realistic natural-language prompts that elicit specified behaviors in frontier open-weight models (Llama 3.1/4, Qwen 2.5, and DeepSeek-V3), using the PRBO to guide the search. These agents automatically discover previously unknown behaviors of concern, including models encouraging self-harm in response to ordinary-seeming prompts.

Note: the image below contains graphic content related to self-harm.

Diagnosing a performance regression on Terminal-Bench with Docent

GPT-5.1 Codex underperforms GPT-5 Codex by 6.5 percentage points on Terminal-Bench, even with the same scaffold. What explains the regression? Using Docent, our AI-backed analysis tool, we find that the regression stems from time-out errors — not performance. GPT-5.1 Codex times out at nearly twice the rate of GPT-5 Codex (49% vs. 29%), despite Terminal-Bench's system prompt never mentioning the time constraint.

Predictive Concept Decoders (PCDs)

We train interpretability assistants to look inside AI models and accurately answer questions about their behavior. These assistants — Predictive Concept Decoders — learn to translate a model's internal states into short, human-readable concept lists, then use those concepts to answer questions about the model's behavior.

PCDs become both more accurate and more interpretable as we scale up training, and they can surface information that models hide or fail to report, including jailbreaks, secret hints, and implanted concepts.