AI

Self-Generated Feedback Destabilizes Test-Time Training: A Causal Decomposition of Long-Horizon Adaptation

Read full article at source ↗

Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 06/10/2026, 09:18:24 UTC.

Briefings

Research published October 2026 found that self-generated feedback during test-time training destabilizes long-horizon adaptation, causing significant prediction errors on independent human text. The study analyzed 128K-token streams across three model configurations (125M, 760M, and 3B) and Adam updates, demonstrating that retaining generated-text updates worsens performance compared to fixed generation or recorded replay methods.

Key points

• Fixed Generation removes over 98% of the damage at 125M and 760M model sizes by using a frozen model to generate training chunks.
• Settlement evaluation shows mean endpoint gaps of 0.07 and -0.02 nats at 125M and 760M while retaining real-text adaptation.

Context

This study investigates how models adapt to long-horizon tasks when trained on their own outputs versus external data, revealing that self-generated feedback introduces instability during test-time training.

Limitations

Results are based on 128K-token streams with snapshot limitations; uncertainty exists regarding the exact causal pathways and whether the observed gaps persist under different conditions.

Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 19/10/2026, 15:07:33 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.

View archived briefing on IPFS
bafybeiflgeyxkfainmtzhv75gxb5whq3qrntqdat7hlzllzi4oegqkwaiy