Why are documents read but not cited harder distractors than random samples?
This explores why 'hard negatives' built from papers an author actually read but chose not to cite teach a model more than randomly drawn papers — and why that same property makes them a tougher discrimination problem.
This explores why documents that were read-but-not-cited make harder training distractors than random samples. The short version: a random sample is off-topic, so a model can reject it on surface features alone — it learns *topic*, not *judgment*. A read-but-not-cited document is on-topic by construction. Someone with expertise looked at it and decided it didn't belong. To separate it from the cited papers, a model has to learn the thing you actually care about: the fine-grained quality and relevance signal that survives after topical similarity is held constant. That's exactly the signal behind learning 'scientific taste' from community citation feedback, where models trained on citation-matched paper pairs learn to predict impact as a capability distinct from execution skill (Can models learn what makes research worth doing?).
The catch is that the harder negatives are, the more they stress the geometry of the model doing the discriminating. Training dense retrievers with structure-targeted hard negatives consistently *degrades* zero-shot generalization — an 8-40% drop — because in high-dimensional cosine space, sharpening one fine distinction warps the rest of the space (Does training for compositional sensitivity hurt dense retrieval?). So 'read-but-not-cited' negatives are valuable precisely because they're near the decision boundary, but that same nearness is what compressed vector representations struggle to honor. The difficulty isn't a bug in the data; it's the whole point, and it pushes against the limits of the representation.
This is the same wall that shows up in matching tasks: a pooled-cosine recall stage cannot reliably reject 'structural near-misses' — candidates that look right in aggregate but fail on the details — and you have to add a separate verifier that operates on full token-to-token interaction patterns rather than a single squashed vector (Can verification separate structural near-misses from topical matches?). Read-but-not-cited documents are the near-misses of the citation world. Telling them apart is a downstream verification problem, not a similarity problem, which is why throwing them in as negatives is both more informative and more demanding.
There's a learner's twist worth surfacing here, because the difficulty cuts both ways depending on who's judging. The reason on-topic-but-rejected items are hard for a *model* is the same reason they're slippery for *human and LLM evaluators*: when topical content is matched, judgment collapses onto shallow cues. LLM judges fall for authority signals and rich formatting independent of substance (Can LLM judges be fooled by fake credentials and formatting?), and users prefer responses with more citations even when those citations are irrelevant — citation count acts as a decoupled trust heuristic (Do users trust citations more when there are simply more of them?). The expert who read a paper and declined to cite it is doing the hard thing these systems fail at: discriminating on relevance after surface plausibility is already satisfied.
The payoff for using these negatives, when it works, mirrors what selective retrieval buys elsewhere: a system that learns *when* a piece of evidence actually earns its place — modeled as a per-step decision to admit or reject knowledge — eliminates noise from unnecessary inclusions and gains over indiscriminate retrieval (When should language models retrieve external knowledge versus use internal knowledge?). A random negative never teaches that admit/reject judgment because the answer is obvious. A read-but-not-cited negative teaches it because the answer was hard enough that a human had to make a call.
Sources 6 notes
Reinforcement learning trained on 700K citation-matched paper pairs successfully teaches models to predict research impact better than GPT-5.2 and generate higher-impact research ideas. Scientific taste emerges as a community-aligned capability distinct from execution skills.
Adding structure-targeted negatives to dense retrieval training consistently degrades zero-shot performance (8-40% nDCG@10 drop) while only partially improving compositional discrimination. This is a geometric trade-off in high-dimensional cosine spaces, not a tuning problem.
A two-stage pipeline—pooled-cosine recall followed by a small Transformer verifier operating on token-token similarity maps—reliably rejects structural near-misses that MaxSim-style late interaction cannot. The verifier succeeds because it operates on full token interaction patterns rather than compressed vectors.
Research identified four evaluation biases in LLM judges, with authority and beauty biases being semantics-agnostic and trivially exploitable through fake references and formatting—zero-shot attacks requiring no model access or optimization.
Analysis of 24,000 Search Arena interactions shows irrelevant citations boost user preference (β=0.273) nearly as much as relevant citations (β=0.285), indicating citation count functions as a decoupled trust heuristic.
DeepRAG models each reasoning step as a Markov Decision Process where the model learns when to retrieve versus rely on parametric knowledge. The 21.99% improvement comes from better-targeted retrieval and elimination of noise from unnecessary external knowledge.