Papers for
content recommendation teams
Papers whose findings have a practical use for this group, as judged from the abstract. Open a paper to read what it means in practice.
Semantic conditional similarity distinguishes document contributions better
Semantic Conditional Similarity (SCS): A Formal Mathematical Framework
Abstract: Semantic Textual Similarity (STS) operates on document pairs, returning a single scalar per comparison. In a multi-document setting where several sources may have contributed to one or more targets, pairwise STS cannot disentangle each source's marginal semantic contribution from their shared vocabulary. In this paper, Semantic Conditional Similarity (SCS) is introduced as a conditional form of STS that computes similarity between two documents after eliminating semantic elements shared with a third via set-theoretic operations. SCS makes four contributions: 1) it introduces directionality, partitioning pairwise relations into lateral (inter-source) and vertical (source-to-target) classes; 2) conditioning induces role-dependent asymmetry by extending the comparison domain from unordered pairs to ordered, conditioned triples of a target, a reference, and a conditioning document; 3) SCS introduces inter-document conditioning by subtracting shared content from a third document, complementing the intra-document conditioning in recent C-STS work; and 4) the set-theoretic decomposition achieves threshold-free partitioning of shared and residual content, requiring no arbitrary similarity cutoff. Although developed for the $n \to 1$ configuration, the framework extends to $1 \to n$ via vertical role reversal and to $m \to n$ via independent composition of source- and target-side conditioning. With the established $1 \to 1$ case, these span the full spectrum of document cardinality. SCS scores should be interpreted ordinally, as rankings of marginal contributions rather than absolute measurements. The broader implication is that metric space axioms commonly assumed in similarity formalizations constrain expressiveness in role-differentiated multi-document settings, and that extending the domain to conditioned triples can recover information that pairwise-symmetric comparisons suppress.
Fast interactive ranking improves document search using AI feedback
Quantum-Inspired Pairwise Ranking Aggregation via Active LLM Feedback
Abstract: Ranking aggregation from multiple retrievers is a powerful yet under-explored technique in modern retrieval and RAG systems. We introduce LLM-IRA, the first extension of the quantum-inspired interactive ranking aggregation framework to document reranking. By leveraging large language models (LLMs) solely for pairwise judgments on high-disagreement document pairs, our method learns fusion weights in a principled, interactive way. The final ranking is a simple, ultra-fast pointwise linear fusion of scores, no full pairwise matrix or expensive listwise prompting is required. Experiments on the BEIR benchmark and TREC Deep Learning tracks with an open-weight instruction-tuned judge (Llama-3.1-8B-Instruct) show that LLM-IRA matches or numerically exceeds strong static fusion baselines on 5 of 6 evaluated benchmarks. We report two weight-update variants: momentum blending and a full accumulated-history refit, which together reach statistical significance (Wilcoxon signed-rank, p<0.05) over the best unsupervised aggregator on 4 of 6 benchmarks (DL19 under momentum; DL20, DBpedia, and NFCorpus under accumulated refit, which is the stronger of the two configurations overall). On the benchmarks we profiled (DL19, DL20), it already clears the best unsupervised baseline within as few as 5 LLM calls per query, well before exhausting its typical ~25--30 call budget. LLM-IRA also achieves cost-effectiveness comparable to state-of-the-art zero-shot LLM rerankers at an order of magnitude fewer LLM calls, and requires no training data or model fine-tuning.
Bayesian methods improve ranking from limited noisy pairwise comparisons
Alternative Strategies for Active Learning with Bayesian Bradley Terry Model
Abstract: Comparative judgement uses pairwise comparisons to rank items, but recovering accurate rankings from a limited number of judgements depends on both the inference method and the pairs selected. We developed a Bayesian Bradley-Terry model (BBTM) that uses Polya-Gamma data augmentation for efficient Gibbs sampling and quantifies uncertainty in latent abilities and ranks. We propose three active learning strategies based on the knowledge gradient (KG) principle. We define our acquisition functions using Kendall's tau distance, posterior rank entropy, and posterior separation. Approximate one-step updates to posterior moments reduce the computational cost of evaluating candidate pairs. We also introduce Posterior Separation Reliability (PSR), which measures the spread of posterior mean abilities relative to posterior uncertainty. It provides an internal indicator of ranking reliability without requiring a reference ranking. In synthetic experiments covering different item counts, comparison budgets, and noise levels, BBTM generally recovers ranking more accurately than preference-based Bayesian comparative judgement (PBCJ). At low to moderate noise levels, the pairwise entropy pair selection strategy performs best in ranking recovery. However, at the highest noise levels, the KG strategy based on posterior separation produces better point rankings. PSR increases alongside improvements in ranking accuracy as comparisons accumulate. Robustness experiments identify the proportion of informative outcomes needed to meet specified accuracy targets within a fixed comparison budget. In a moderate noise setting, replacing 30\% of outcomes with random choices has little effect on ranking performance. These findings demonstrate the value of Bayesian inference and active learning pair selection for recovering ranks from limited, noisy paired comparisons.
Ranking methods differ in preference and position accuracy in search results
Preference Is Not Position: Evaluating Reranking on NevIR in Fixed Hybrid Candidate Pools
Abstract: Near-identical passages can differ in a decisive condition, such as negation. A ranker may order such a pair correctly while placing the preferred passage far down the retrieved list. We examine this distinction using official NevIR pairs jointly covered by fixed candidate pools from dense, learned sparse, and BM25 retrieval. On 2,736 Test query directions, a pairwise-trained ranker improves strict per-query preference over fixed fusion (56.62% versus 52.30%) while reducing preferred-passage mean reciprocal rank (MRR) from 0.5764 to 0.3055. An exploratory fixed-model replay of a variant with background sampling and revised supervision yields MRR of 0.5859 and strict preference of 54.53%. Bounded pointwise reranking of the fusion top 20 improves both outcomes. Qwen3-14B-AWQ with fusion tie-breaking reaches 77.45% strict preference, 58.99% NevIR-style paired accuracy across 1,363 complete pairs, and preferred-passage MRR of 0.7509. A BGE cross-encoder reaches 69.77% strict preference and MRR of 0.7404. A post-hoc analysis finds that the branch where equal Qwen grades for the two designated passages trigger fusion tie-breaking accounts for 27% of evaluated directions and contains most remaining strict-preference failures. These findings support joint evaluation of pairwise preference, designated-passage position, and computational cost, while distinguishing model judgments from the combined ranking procedure.
Reading position predicts personal highlights better than popularity or similarity
Reading Position Is the Baseline to Beat: A Time-Ordered Evaluation of Personalised Highlight Prediction
Abstract: A reader's first highlights on a page are the cheapest personal signal a reading product has. The natural plan is to suggest what similar earlier readers marked, and to judge the result against popularity. We argue that the baseline to beat is reading position. In a time-ordered evaluation on one social highlighting platform (7,343 reader-page pairs on 1,511 pages after one highlight), ranking the sentences just below a reader's first highlight, with no other reader's data, puts the next highlight in the top five 47% of the time, against 26% for popularity and 29% for the better of two similarity methods. The baseline depends on the target: over all later highlights that ranking loses to popularity, while popularity discounted by distance from the latest highlight, at the scale with the best average precision of three tried, beats popularity and both similarity methods on both targets. In a comparison specified in advance, neither similarity method shows a gain over popularity in average precision over all later highlights, from one to five highlights, and a gain of +0.01 is excluded. Nor would a gain by itself show that a method has found a reader's preferences: synthetic readers who share one set of preferences produce one, and an evaluation out of time order shows a method where the reader went. The position results are exploratory and unconfirmed. Personalisation inside a document should be evaluated in time order and against reading position.
Humor ranking wins contest using multi-stage AI retrieval system
IROH: Insightful Ranking Of Humor using Multi-Stage Hybrid Retrieval with Rationale-Distilled LLM Judges for JOKER 2026 Track Task 1 English
Abstract: Our team, VANGUARD, presents IROH (Insightful Ranking of Humor), a three-stage retrieval system for JOKER Task 1 English at CLEF 2026, achieving first place on the leaderboard with 0.6347 MAP. Our pipeline combines hybrid sparse-dense retrieval, cross-encoder reranking, and a LoRA-adapted Large Language Model judge ensemble. We employ Gemma 4 to generate query-aware rationales under two prompt strategies, generic and typed, and produce up to four types of structured hard negatives for training data construction. Through an ablation across three cross-encoder architectures, four dense embedders, and eight judge configurations, our key findings are threefold: (1) the rationale-distilled judge is the primary driver of ranking quality, whereas appending rationales to the first-stage index contributes negligibly; (2) structured hard negatives degrade generalisation in nearly all configurations despite inflating local validation scores; and (3) across the components we ablate, the lighter, better-calibrated model is competitive with or stronger than its larger counterpart, with the generic-rationale Qwen2.5-7B judge (0.6055 MAP) outperforming every Gemma-4-31B configuration, and the advantage of generic over typed rationales is concentrated almost entirely in the smaller model.
Compact binary codes improve document search efficiency and accuracy
Matryoshka Hash Representations for Model-Aware Compact Semantic Retrieval
Abstract: Retrieval-augmented generation (RAG) depends on dense retrieval: each document is stored as a learned vector, and a query is answered by finding its nearest neighbors in that vector space. Keeping one full-precision vector per document is the dominant index cost at corpus scale, so retrieval systems replace each vector with a short code of a few bytes---a step called quantization. Standard quantizers such as product quantization (PQ) pick the code that reconstructs the original vector most closely. A single code is even more useful if it serves several byte budgets at once: when its short prefixes are each directly searchable, a deployment can set its efficiency--quality operating point without re-encoding the corpus. But training all prefixes under one objective makes the early bits a compromise across budgets---short codes improve while the full-width code degrades. Quantization to low-bit representation, such as binary codes, further sharpens the conflict. We introduce Matryoshka Hash Representations (MHR), a two-stage procedure that separates full-width training from prefix organization. MHR first learns a longer binary code, then freezes the model and trains additional zero-initialized residual code adaptors for directly searchable prefixes. Documents are stored at one bit per coordinate, while queries keep continuous logits like PQ to attain sufficient expressivity. We implement the search process with FAISS FastScan. Trained on MS MARCO and zero-shot transferred to seven BEIR datasets, MHR reaches .5561 NDCG@10 and .6535 Recall@100 at 32 bytes, surpassing the best baseline of the same budget. The advantage is more pronounced in lower budgets. The same code also strengthens two common pipelines: shortlisting candidates for full-precision reranking, and pruning a low-storage graph index such as LEANN.