The author highlighted it. You dropped it.

Three papers about the layer underneath the model: what your document converter silently throws away, what your stacked retrieval upgrades actually add up to, and what a citation check is really checking.

Today's pick

0.690 → 0.150

micro F1 for recovering the requirements an author flagged, formatting-aware ingestion against a mainstream converter on the same files

What the Author Wrote, What the Agent Reads: Silent Cue Loss in Enterprise Document Ingestion

An author writes a policy document and highlights the three clauses that are actually binding; your pipeline converts the file to text, the highlighting is gone, and the agent now reads sixty sentences of equal weight and answers confidently about the wrong ones. Scored against each file's own stored formatting by a deterministic string matcher rather than a model judge, recovering the flagged requirements from 76 real policy documents runs at 0.690 micro F1 when ingestion preserves how the text looked, 0.150 through a mainstream converter and 0.075 through plain-text extraction, and the converter's false-positive rate on clauses the author did not flag is four times higher. Drop each cue in turn and the entire effect is two of them — highlight is worth 0.302 and font colour 0.262, while bold, italic, monospace and strikethrough all land within 0.003 of nothing. Pulling those two out of the file is an afternoon's work in your extractor, saying in words what the cue means adds a further 0.085 on top of merely keeping it, and a third of 6,462 real documents carry one.

open the explainer → abstract PDF

+47% → +17%

retrieval gain predicted by adding up three techniques measured alone, against what the three delivered once combined

Retrieval Augmented Generation Compositionality

You read that a cross-encoder reranker bought +X on a benchmark table and a document-aware prefix bought +Y, so you add both and budget for roughly X plus Y. Across 450 configurations over eight retrieval and answering benchmarks, with corpora from 5,000 to 200,000 documents, three techniques whose isolated measurements predict a 46.6% ranking improvement deliver 17.2% together — a 63% shortfall that grows to 72% once all five individually-helpful techniques are stacked. Worse for planning, ranking quality and answer faithfulness are moved by different halves of the pipeline: the embedding and index-enrichment choices that dominate ranking barely touch whether the answer sticks to its sources, while query rewriting and agentic verification dominate faithfulness and do nothing for ranking, and one of those verification gates stripped 14 of 15 supporting passages in a multi-hop case. Tune the whole stack at the corpus size you actually serve — the configuration that wins at 5,000 documents trails by 13.7% at 200,000, and the crossover is already visible by 25,000.

open the explainer → abstract PDF code

3 of 327

citations actually about the claim, out of the 327 real database records a model answering from memory produced

Real PMIDs, Wrong Papers: Harness Design for Biomedical Claim-Verification Agents

If your product cites its sources, you probably guard it by checking that the identifier resolves. Asked to verify 100 scientific claims from memory, a model emitted 335 reference identifiers, 327 of which are genuine records in the live literature database — and a judge from a different model family rated 3 of those 327 as actually about the claim. An existence check therefore waves through 98% of invented citations, while the same audit rates 67% of a search agent's citations relevant, so the guardrail you want compares the cited text to the claim rather than the identifier to the index. The rest of the harness comparison points the same way: handing the raw claim to the search engine finds the right paper for 6% of claims and letting the model write its own queries roughly triples that, verdicts land at 62% against 79-86% when the correct sources are supplied outright, and the agents used only 1.42 to 1.70 of their four allowed searches — the bottleneck is retrieval, and raising the budget alone will not move it.

open the explainer → abstract PDF code