LLMs show lost-in-the-middle failures on 100000-token electronic health records
Researchers document that electronic health records now routinely exceed 100,000 tokens per patient, a scale where large language models show lost-in-the-middle behavior: center context is retrieved less reliably than information near the edges. An arXiv paper characterizes this failure in clinical long-context reasoning and proposes inhibitory attention as a mitigation, stressing that missing mid-record labs or history notes is clinically non-benign because those middle details often drive decisions. For operators building EHR ingestion and prompting pipelines, usable long context depends on position-dependent retrieval quality, not token budget alone. The arXiv identifier and v1 announcement timing should be verified before treating empirical validation and baselines as settled.
LLMs show lost-in-the-middle failures on 100000-token electronic health records
Researchers report electronic health records now routinely exceed 100,000 tokens per patient and that large language models retrieve center context less reliably than edges. The arXiv paper proposes inhibitory attention to mitigate lost-in-the-middle effects in clinical long-context reasoning.
Key takeaway
In clinical EHR workloads exceeding 100,000 tokens, lost-in-the-middle retrieval bias turns long context from a capacity win into a safety hazard for mid-record evidence.
What happened
According to reporting in an arXiv cs.CL preprint, electronic health records now routinely exceed 100,000 tokens per patient, yet large language models retrieve information near the center of those long contexts less reliably than details at the edges.
The paper characterizes this lost-in-the-middle effect in clinical long-context reasoning, proposes inhibitory attention as a mitigation method, and frames the bias as clinically non-benign because middle clinical details drive decisions.
Evidence
Electronic health records now routinely exceed 100,000 tokens per patient
arXiv cs.CL · attributed
Electronic health records now routinely exceed 100,000 tokens per patient
Large language models retrieve center context less reliably than edge information in long EHR inputs
arXiv cs.CL · attributed
information near the center of a long context is retrieved less reliably than information near the edges
Researchers propose inhibitory attention to mitigate lost-in-the-middle effects in clinical long-context reasoning
arXiv cs.CL · attributed
They propose 'inhibitory attention' to mitigate this
Lost-in-the-middle failure in clinical use is framed as non-benign because middle details drive decisions
arXiv cs.CL · attributed
the failure is clinically non-benign because middle clinical details drive decisions
Why it matters
Builders deploying LLMs on long clinical inputs need to account for position-dependent retrieval failure, not just total token count, when designing EHR ingestion and prompting pipelines.
Limits and uncertainties
The arXiv ID (2608.20348) and v1 announcement appear inconsistent with a real publication, so the method's empirical validation and baselines should be verified before trusting results.
Practical implications
Design EHR ingestion and prompting pipelines to test retrieval by context position, not only whether the full record fits within the model window.
What to watch
Independent verification of the arXiv preprint identifier and replication of inhibitory-attention baselines on representative 100,000-token EHR lengths.
Original reporting: Inhibitory Attention for Clinical Long-Context Reasoning: Characterizing and Mitigating Lost-in-the-Middle Effects in EHR Processing
