long-context-tasks-over-8k-128k-token-documents-middle-of-context-penalty-granularity-dependent-word-level CLM-0429In long-context tasks over 8K-128K-token documents, the middle-of-context penalty is granularity-dependent: word-level QA and paragraph-filling accuracy dip when the evidence sits mid-document, but sentence-cloze accuracy peaks in the middle, because mid-paragraph sentences carry concrete entity and date overlap that models match on instead of reasoning about discourse role.
Capability: Losing information in long inputs
Observed on
12 long-context LLMs, zero-shot, UN reports 8K-128K tokens, six official UN languages, position stratified as begin/middle/end at paragraph and document level..
Sources
- Measured across 12 models on a newly built benchmark with position-stratified items; the mid-peak explanation is supported by LLM- and human-annotated error analysis of cloze cases where over half of models failed, not by a controlled manipulation of surface overlap. Self-built benchmark, no independent replication.
Disagreeing is the most useful thing you can do here. Nothing on this page waits for a reviewer: the counts above move when evidence lands, and every source stays attached with who brought it.
Related claims
- On long-context tasks over 8K-128K-token documents in the six official UN languages, models answer word-level span questions well (about 73% average accuracy across 12 models) but degrade sharply as the required unit coarsens to sentence selection (about 21%), paragraph filling and document-level summarization and translation, so long-context competence measured by needle-style word retrieval overstates comprehension of larger discourse units.Losing information in long inputs · reviewed by AI
- In ESG-report RAG question answering with retrieval held fixed at a single top-1 chunk, open-weight 2B-30B generators vary widely in how well they stay anchored to the retrieved passage (faithfulness 0.607-0.822) yet all land in a narrow low band of factual correctness (0.387-0.449), and the most faithful model is the least factually correct, so groundedness to context and correctness of extracted ESG facts do not track each other.Competence in a specialist domain · reviewed by AI
- For multi-hop QA over 28k-896k token haystacks with frozen Qwen2.5 backbones (7B-72B), summarizing the document in a parallel hierarchical tree — overlapping windows scanned ~5x, raw segments scored 'critical' passed straight to the final answer, and one summarization channel per decomposed sub-question — scores higher than sequential memory-agent and BM25 retrieval baselines without training, and beats an RL-trained memory agent on the out-of-distribution dataset while losing to it in-distribution.Losing information in long inputs · reviewed by AI
- When claim-level uncertainty is estimated from a response-claim graph, anchoring the graph with trusted external priors (edges weighted by entailment from verified source sections) and scoring claims by weighted closeness centrality separates ungrounded claims from repeatedly-generated-but-false ones better than internal-consistency centrality alone, and yields retrieval whose answerability tracks its accuracy instead of the near-100% answer rate GraphRAG shows on the same long documents.Stating false facts confidently · reviewed by AI
- On text-to-SQL questions that are underspecified with respect to the schema, retrieving clause-level fragments (conditions, joins, calculations, groupings, output specs) from a database's own past successful SQL queries and prepending them to existing few-shot pipelines raises execution accuracy across four different pipelines, but the gain saturates and can reverse once too many candidates are retrieved, because high recall with low precision lets the generator pick a plausible-but-wrong pattern.Writing a reusable locator: regex, SQL, XPath · reviewed by AI
Notes
Drafted from the paper by a model and filed unreviewed. Visible here so it can be read, not because anyone has vouched for it: it does not move any technique's standing and does not count toward the internal scorecard. Drafted confidence: medium. Falsifier as drafted: A replication stratifying by granularity that finds sentence-level cloze also dips in the middle, or that the mid-position peak survives when surface lexical overlap between candidate sentences is equalized across positions. Drafted stance toward performance-multi-document-qa-highest-when-answer-at-start: contests -- The paper finds the middle-position dip holds for word and paragraph tasks but reverses for sentence-level cloze, so the positional curve is not uniform across task granularity.