DyMT-ESB Paper: LLM Social Bias Emerges Late, Fluctuates and Re-emerges in Multi-Turn Chat
Bollegala · x · 2026-09-17
The paper "DyMT-ESB: Dynamic Multi-Turn Evaluation of Social Bias in User-LLM Interactions" (EMNLP 2026 Findings) introduces a protocol beyond fixed-turn, pre-scripted evals: follow-up user queries are generated from the evolving dialogue history, allowing variable-length evaluation.
Key findings:
- LLMs exhibit social bias even in coherent, response-conditioned multi-turn interactions;
- Bias can emerge late in a conversation, fluctuate across turns, and re-emerge after subsiding.
The authors argue social bias should be evaluated as a turn-level dynamic phenomenon, exposing blind spots in template-based, fixed-length benchmarks.
More from Safety
- AI code's security bugs are rarely bad code -- they're missing code — RyzeBlaziken · 2026-09-17
- n8n hit by CVSS 10.0 chain: unauthenticated file read to full RCE, PoC out — evilsocket · 2026-09-17
- Devs ask: how to actually stop agents before they do damage, not just prompt them — Real_KingZeotic · 2026-09-17
- King Charles Warns AI Poses Existential Threat, Calls for Control 'Before It Is All Too Late' — ShakeelHashim · 2026-09-17
- Hands-on test of Typesafe.ai's new Jev model shows promise for secret detection — teyhouse · 2026-09-17
- Hugging Face was first to spot, remediate and disclose the rogue agent attack, says ex-HF lead — TheZachMueller · 2026-09-17