Steering Qwen3-8B with the <|im_end|> vector cuts verbose replies dramatically
Sauers_ · x · 2026-11-19
A neat activation-steering demo: adding the unembedding vector of the end-of-turn token <|imend|> to layer 18 of Qwen3-8B's residual stream drastically curbs the model's filler output. Per the shared chart, an unsteered 100-token reply drops to 15 tokens at strength 1, and 29 of 30 replies end immediately at strength 4 (p=0.02), while random-direction steering has the opposite effect, inflating replies to a 200-token cap. A lightweight, tuning-free way to control response verbosity.
More from coding & agent
- ClaudeAI Weekly: Single-HTML-File Arms Race, 8.9B-Token Strawberry, Post-IPO Anthropic — ClaudeAI-mod-bot · 2026-11-22
- Agno's AgentOS turns any agent into a production FastAPI app with MCP, RBAC and tracing — pritisinghhhh · 2026-11-22
- Chinese devs vent: app store rejections and copycats are just the norm — tinyfool · 2026-11-22
- Compacting Early and Often Boosts Agent Accuracy While Cutting Cost and Runtime — zainhas · 2026-11-22
- Astra 6.2 silently skips modules in audit tasks, then claims it checked everything — MbahSurip · 2026-11-22
- Is Haskell 'unsafe' too? A debate on types vs verification in the LLM era — satnam6502 · 2026-11-22