AI SEARCH · 2026-10-25
AI Search — Paper Daily, 2026-10-25
Adobe splits the agent-retrieval attribution gap, a conversational simulator ranks like NIST, marginal viewpoint coverage and cross-query RCP-nDCG arrive—today’s AI Search thread is that eval and evidence selection must admit the generation layer consumes retrieval before the user, and replace click-SERP assumptions with calibrated, viewpoint-preserving objectives.
4.60
LEAD STORY ·
AdobeAdobe CTR, MRR and behavioral NDCG assume the user sees and acts on ranked results. In an agentic assistant a generation layer consumes retrieval first, so attribution breaks on who saw the results, when the user acted, and what caused the outcome—Adobe calls this the attribution gap. LARA (Layered Agentic Retrieval Assessment) measures retrieval quality, engagement and downstream outcomes by layer; production retrieval rate on assistant prompts is only 0.31%, and chat users reformulate more than gallery users (44.4% vs 33.1%). Do not evaluate agent retrieval as if the user clicked a SERP—attribute by layer.
4.35
Conversational search needs trustworthy user simulators for scale. AutoSim’s relevance judgments rank a controlled pool of retrieval systems in agreement with NIST across three TREC iKAT years (Kendall τ 0.66 / 0.91 / 0.91, all p≤0.002), and that agreement survives removing its persona. Self-improving against held-out real-user signal, weight-level self-improvement drives real-vs-sim AUC from ~0.85 down to 0.45, lifts retrieval +0.136 nDCG, and returns a relevant passage on 57% of turns. A simulator is trustworthy when it ranks systems like NIST, not when it merely sounds human.
4.35
Multi-viewpoint RAG must keep distinct answers and stances under a fixed context budget. LMVU ranks by the marginal viewpoint coverage a candidate adds to the current selection, trained with viewpoint annotations that compare candidates in the same selection state, and rescoring budget-feasible candidates as context grows. It reaches 94.6% coverage on Perspectrum and matches or nears full coverage on WVS/ESS10, beating the strongest baselines. For multi-viewpoint retrieval, rank by marginal coverage, not per-doc relevance alone.
4.35
Embedding and generation are usually separate specialists with incompatible objectives; unified models like GritLM either switch attention regimes or train the two as parallel losses. ULME proposes predictive coupling so embedding and language modeling reinforce each other—aimed at RAG/agent pipelines that embed then generate the same query. Unifying embed+generate is about objective coupling, not shared weights plus a mode switch.
4.35
CCohere
As rerankers cluster in quality, nDCG on sparse discrete human labels stops separating them. LLM relative judgments split close docs but share no scale across queries; absolute grades share a scale but are too coarse. Cohere’s Rubric-Calibrated Preferences (RCP) uses IRT for cross-query calibration; RCP-nDCG replaces discrete labels: useful-vs-not AUC 0.910 (vs 0.651 for benchmark labels); among 185 comparisons where only one metric agrees, RCP-nDCG is that metric 72.4% of the time; on TREC-DL it sides with NIST grades whenever they separate. nDCG failure modes hit 45.6% of NanoBEIR and 30.7% of BRIGHT queries. Reranker eval must move from discrete nDCG to dense, cross-query-calibrated judgments.
4.20
Search agents often lack search-state awareness—what is already obtained vs still missing—so they hallucinate reasoning and stop early. SARL (Self-Assessment Reinforcement Learning) trains models to assess their own search state and decide whether to keep retrieving or answer. Stop conditions should come from self-assessed information gaps, not step budgets or a single hit.
4.00
KG-RAG often distills continuous text into discrete symbols (semantic loss) and then clashes fragmented graphs with the continuous text LLMs were pretrained on. This paper couples graph topology with energy-driven pathfinding to preserve contextual consistency in RAG and ease structuring loss and modality mismatch. KG-RAG bottlenecks are often how knowledge is structured and traversed, not only retriever scores.
4.00
CLIP’s 77-token text encoder truncates long captions. ERCLIP is a small-data adaptation: at train time long captions are sentence-encoded then fused by mean pool or a gated residual Transformer scorer with their own contrastive loss; at inference it is an unmodified dual encoder still reading the 77-token window. It gains 6.5–11.2 points over base CLIP and closes 62–77% of the gap to Long-CLIP. Long captions need not change the inference encoder—inject sentence-level signal at train time.
3.90
Visual document retrieval needs rich multi-vector page evidence, so encode/match is expensive. Compression and elastic reps expose many budgets but do not say which budget a query should use. FlexVDR learns to estimate retrieval value of alternative Matryoshka representation budgets before paying full compute, then allocates per query. Elastic representations matter when you allocate budget per query, not when you only expose configs.
3.90
OCR-free VLM visual document retrieval emits too many tokens for scale; heuristic compressors struggle to balance importance, semantic redundancy and coverage. This paper casts visual token compression as unbalanced optimal transport (UOT), unifying those three under one objective while keeping retrieval quality. Spend VDR token budget as a transport problem once, not as a pile of heuristics.
EMNLP 2026
ECCV 2026