Lost in the Middle: LLMs Miss Facts Buried in Long Contexts, Study Shows
thisdudelikesAI · x · 2026-09-12
The final entry in a 12-paper series for AI engineers: Liu et al. (2023), Lost in the Middle (TACL 2023).
Key findings:
- Tested on multi-document QA and key-value retrieval, models don't robustly use long input contexts
- Performance peaks when relevant info sits at the beginning or end of context, and degrades significantly in the middle — even for explicitly long-context models
- Where you place a fact changes whether the model uses it
The author calls it the most practical paper for anyone stuffing giant prompts; the study also introduced new evaluation protocols for long-context models. The prior entry in the thread recommended DSPy: treat prompting as a program to optimize, not a string to tweak.
More from Research
- Nature MI paper unifies neural superposition and sparse interpretable codes in one framework — GretaTuckute · 2026-09-12
- Astribot's SmoothRL: online RL learns only from actions the robot actually executes — jiqizhixin · 2026-09-12
- Building adebench: scoring what your agent's memory actually delivers, 93.7/100 — Soft-Lie-434 · 2026-09-12
- ECCV debates: are explicit 3D representations obsolete as video generative models take over? — qixing_huang · 2026-09-12
- OpenAI claims a Millennium Prize problem, and mathematicians are uneasy — The Verge AI · 2026-09-12
- Fruit fly brain sim wired to a 1B LLM: 139,255 neurons now chat with you — TinfoilTricorn · 2026-09-12