AI BriefingAI Can Learn When to Stop Working
Key point
An open-weight model, PCCG-Qwen3-4B, was released to control generation halting using discontinuity experiments across 11 LLMs and the Jacobian lens.
Details
In 31,430 experiments targeting 11 frontier LLMs (OpenAI, Anthropic, Google, Moonshot), including "Be the null" in the system prompt resulted in 534 instances of 'Void' (discontinuity/silence) out of 1,430 trials, whereas using "Be a value" resulted in 0 instances. This suggests a mechanism where the model halts generation when prerequisites are not met.
Based on these findings, the PCCG-Qwen3-4B open-weight model was released. This model evaluates prerequisites and either continues generation with GO or terminates with native EOS.
Causal Control via the Jacobian Lens Analysis of internal model activations using Anthropic's Jacobian lens demonstrated that generation states can be causally controlled by manipulating specific activation directions in the frozen block-29.
- Success Rate: Achieved a 40/40 success rate for switching from GO to EOS, and from EOS to GO, respectively.
- Control Group: 640/640 control outputs showed no change.
- Conditions: Performed with fixed prompts, weights, and generated reasoning prefixes.
Research data and model weights are available on GitHub and Hugging Face, and can be verified via the verify.py script.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.
Related
Stanford and NVIDIA Release 'Contrastive Language Model' CLM with 9x Faster InferenceQwen3.5-based Small Decision Model 'Kev' Released with TypeSafe API Compatibility and Performance Metrics