Can reranking candidate summaries improve perspective representation better than prompting?
This explores whether selecting among already-generated candidate summaries (reranking) is a stronger lever for capturing varied perspectives than just instructing a model to do so in the prompt — and the corpus suggests the answer leans yes, because prompting has a hard ceiling that selection-based methods don't.
This reads the question as a contest between two control surfaces: shaping a single summary through prompt instructions, versus generating candidates and reranking them toward an explicit objective. The corpus doesn't have a paper squarely on 'perspective representation,' but it has strong material on both sides of the underlying mechanism — and the lateral picture favors reranking.
The case against relying on prompting is the most direct finding here. Prompt optimization can only reorganize what a model already holds; it cannot inject what's missing, which creates a hard ceiling no clever instruction can climb past Can prompt optimization teach models knowledge they lack?. Worse, when a model's training priors are strong, textual prompting alone fails to override them — the model keeps generating what its parameters 'want' regardless of what you ask Why do language models ignore information in their context?. For perspective coverage specifically, that's exactly the failure you'd fear: a model defaults to the dominant, well-represented viewpoint and a prompt asking it to 'represent multiple perspectives' bounces off the prior.
Reranking sidesteps this because it operates on a pool of already-produced candidates and optimizes selection toward an actual target rather than hoping the generation step complied. Two notes show this works. ReLSum trains a summarizer using the downstream relevance score as the reward, producing summaries aligned to what the ranking task actually needs instead of generically fluent prose — and it beats generic summaries on the real metric Can reinforcement learning align summarization with ranking goals?. METEORA goes further and is the sharpest doorway here: rationale-driven selection — picking content by explicit reasons rather than surface similarity — beats similarity re-ranking by 33% while using half the chunks Can rationale-driven selection beat similarity re-ranking for evidence?. That last result reframes the question: not just 'rerank vs. prompt,' but 'rerank on what basis' — selecting candidates for *why* they add a perspective beats selecting them for how similar they look.
The broader pattern across the corpus is that explicitly modeling the objective beats hoping it emerges. Ranking systems that don't explicitly model selection bias collapse into degenerate equilibria that amplify their own dominant past choices Why do ranking systems need to model selection bias explicitly? — a structural echo of how a prompted model collapses toward its dominant prior. And where models lack a behavior entirely, the fix is a training signal, not an instruction: topic resilience comes from fine-tuning on distractor dialogues, because models learn what-to-do but not what-to-ignore from prompts alone Why do language models engage with conversational distractors?.
So the synthesis: reranking candidate summaries should improve perspective representation more reliably than prompting, but for a non-obvious reason — it's not that reranking is magic, it's that prompting is structurally blocked from overriding strong priors and injecting absent coverage, while a generate-then-select loop optimized toward an explicit representation objective (especially a rationale-based one) can. The thing worth knowing you wanted to know: the biggest gains may come not from reranking itself but from *what you rank on* — reasons over similarity.
Sources 6 notes
Prompting works entirely within a model's pre-existing training distribution and cannot supply domain knowledge absent from training data. This creates a hard ceiling: no prompt strategy can compensate for missing foundational knowledge, only reorganize what already exists.
Research demonstrates that LMs generate outputs inconsistent with their context because parametric knowledge from training dominates over in-context information. Textual prompting alone cannot override strong priors; causal intervention in representations is required.
ReLSum trains summarizers using downstream relevance scores as RL rewards, producing dense, attribute-focused summaries instead of fluent prose. This alignment to the actual ranking metric improves recall, NDCG, and user engagement in production e-commerce search.
METEORA uses LLM-generated rationales with flagging instructions to select evidence, achieving 33% better accuracy with 50% fewer chunks than similarity re-ranking across legal, financial, and academic domains. The method also improves adversarial robustness substantially.
YouTube's multi-objective ranker uses MMoE for conflicting objectives and a shallow position tower to remove selection bias from training data. Without both mechanisms, models converge on degenerate equilibria that amplify their own past decisions.
Fine-tuning on just 1,080 synthetic dialogues with distractor turns significantly improves topic resilience, revealing that the gap is not model capacity but absent training signal. Models learn to follow what-to-do instructions but not what-to-ignore instructions.