Reproducing the J-Space Hallucination Signal

dasjomsyeet · reddit · 2026-07-12

The author stress-tested Anthropic's J-Space / workspace noise hallucination detection signal on Qwen3-4B across 7 datasets and roughly 11,400 samples to see if it could serve as a deployable error router.

Main Findings

Great at catching "high-confidence but wrong" factual questions: On long-tail factual retrieval tasks like PopQA, workspace noise is better at identifying errors than just looking at output logprobs; with a 5% manual review budget, it even achieved 100% precision.
Almost ineffective for "internalized fallacies": On TruthfulQA, the metric's predictive power virtually fails; even when the model is in a "high output confidence + low workspace noise" zone, it still answers incorrectly in many cases.
Math tasks break static thresholds: Directly migrating a threshold calibrated on TriviaQA to GSM8K fails because mathematical reasoning inherently brings higher noise scores.

Conclusion

The author believes this signal is more about identifying when "the model is struggling to piece together a plausible-sounding fake answer" rather than whether the model actually grasps the facts. Thresholds cannot be universally applied across different task types. Current results are based on a single model and a single evaluation run; next steps involve sweeping parameter scales. The repository provides raw data, metrics, confusion matrices, and a Colab notebook.

Original post →

Reproducing the J-Space Hallucination Signal

More from Models