IR 2 — Paper menu

How to choose. Read the abstracts of the papers that interest you (the bidding page shows each one in full), then rank 3 to 5 of them there by Sept 30: name, email, whether you took the spring IR course, an optional preferred partner. Your bids must cover at least three different sessions, with at least two on Nov 13 or Nov 20; presenting early keeps December free for other exams. Bids are visible to everyone and can be changed until the deadline. On Nov 4 (postponed from Oct 5) pairs are formed (a newcomer with a returning student) and papers assigned to maximise everyone's preferences, one paper per pair; discussant roles are published the same day.

What you commit to. A 25-minute talk, a 10-minute live demo (three queries: hero, failure, wildcard), a reproduction of one key result at small scale, and a run of the method on the shared testbed: Track A (SciFact, NFCorpus, ArguAna and the scientific-article set, nDCG@10 with a p-value) for ranking papers, Track B (the 92k MS MARCO subset, queries per second, index size, RAM) for efficiency papers. See program §6 for the testbed, the compute guide and the contrastive-training starter notebook that papers marked training build on. You present once, with your partner, and you are discussant once for another pair's paper: you read it and their code, send them two questions in advance, and lead the dialogue in the session. Difficulty ★ = accessible, ★★ = solid work, ★★★ = ambitious; compute laptop / Colab / Colab + a few euros of API. The full annotated review with more detail per paper is full review.

Every paper below has public code or models unless stated. "Classic" = the reference everybody cites; "recent" = 2024–2026.

Session 2 (Nov 13) — Representations: how embeddings are trained, and where they break

1. SimCSE: Simple Contrastive Learning of Sentence Embeddings (Gao, Yao & Chen, EMNLP 2021) · classic · ★ · laptop/Colab · training

https://arxiv.org/abs/2104.08821 · code: github.com/princeton-nlp/SimCSE
Two dropout-perturbed passes of the same sentence are a positive pair; in-batch negatives do the rest. The paper explains its gains with the alignment/uniformity geometry of the embedding space.
You: train unsupervised SimCSE on the testbed abstracts with MiniLM/BGE-small, then replace dropout by "title vs abstract" views of the same paper and measure retrieval, alignment and uniformity. Does the instructor's patent finding (dropout over-disperses) hold on scientific text?

2. LLM2Vec: Large Language Models Are Secretly Powerful Text Encoders (BehnamGhader et al., COLM 2024) · recent · ★★ · Colab · training

https://arxiv.org/abs/2404.05961 · code: github.com/McGill-NLP/llm2vec
Three cheap steps turn a decoder LLM into an encoder: bidirectional attention, masked next-token prediction, unsupervised SimCSE. The counter-paper Seq vs Seq / Ettin (Weller et al., ICLR 2026) says purpose-built encoders still win at equal size.
You: apply LLM2Vec to Qwen2.5-0.5B/1.5B, ablate the three steps on SciFact/NFCorpus, and compare with an equally sized encoder (ModernBERT-base) at the same budget.

3. Contextual Document Embeddings (Morris & Rush, ICLR 2025) · recent · ★★ · Colab · training

https://arxiv.org/abs/2410.02525 · code: github.com/jxmorris12/cde
A document embedding should depend on its corpus, like BM25's IDF does. Contextual batching (cluster the corpus, make in-batch negatives hard) plus a two-stage encoder conditioned on a corpus sample.
You: evaluate cde-small-v1 with context from the right corpus, the wrong corpus, and none; reproduce the clustered-batching gain when fine-tuning BGE-small. Add late chunking (Jina 2024) as the within-document cousin.

4. On the Theoretical Limitations of Embedding-Based Retrieval (LIMIT) (Weller et al., ICLR 2026) · recent · ★★ · laptop

https://arxiv.org/abs/2508.21038 · code: github.com/google-deepmind/limit
A single vector of dimension d can only express a bounded number of top-k subsets. The LIMIT dataset ("who likes apples?") breaks state-of-the-art embedders while BM25 and multi-vector models cope. A 2026 rebuttal (R^2k is theoretically large enough, ICML 2026) disputes the bound.
You: reproduce the free-embedding experiment on SciFact/NFCorpus qrels (smallest d that fits them), run BGE-small, Qwen3-Embedding-0.6B, BM25 and a MaxSim scorer on LIMIT-small, and take a side in the debate.

5. SPLADE: Sparse Lexical and Expansion Model for First-Stage Ranking (Formal et al., SIGIR 2021; SPLADE-v3 2024) · classic · ★ · laptop/Colab · training · Track B

https://arxiv.org/abs/2107.05720 · https://arxiv.org/abs/2403.06789 · code: github.com/naver/splade, model naver/splade-v3
BERT's masked-LM head projects each token onto the vocabulary; a sparsity regulariser keeps the result usable in an inverted index. The neural successor of BM25, and the middle ground between BM25 and dense retrieval.
You: reproduce SPLADE-v3 on the testbed with sentence-transformers' SparseEncoder, then fine-tune a small SPLADE on the MS MARCO subset, sweep the FLOPS regulariser, and inspect the expansion terms on scientific text.

6. Timeline talk: static embeddings are back — Model2Vec and static retrieval models (Aarsen, HF blog 2025; Model2Vec 2024) · recent, non-peer-reviewed · ★ · laptop

https://huggingface.co/blog/static-embeddings · github.com/MinishLab/model2vec · background: Zhang et al. 2025 survey https://arxiv.org/abs/2507.20783
Word2vec-style static vectors, trained with 2025 recipes, reach ~85–90 % of a sentence transformer at 400× the speed. The talk is a history of embeddings 2013→2026: word2vec → SBERT → contrastive pretraining → LLM embedders → static again.
You: "one model per era on one testbed": word2vec averaging, MiniLM (2019), Contriever (2021), E5/BGE (2022–23), Qwen3-Embedding-0.6B (2025), static-retrieval-mrl (2025); plot nDCG@10 against release year and CPU latency. Then distil BGE-small into a static model yourself.

Session 3 (Nov 20) — Domains: scientific papers, patents, adaptation, languages

7. SemCSE: Semantic Contrastive Sentence Embeddings Using LLM-Generated Summaries for Scientific Abstracts (Brinner & Zarriess, 2025) · recent · ★★ · Colab · training

https://arxiv.org/abs/2507.13105 · code: github.com/inas-argumentation/SemCSE
Instead of citations, an LLM writes several summaries of each abstract; summaries of the same paper are positives. Second-best on SciRepEval without any citation data. Compare with SciNCL (Ostendorff et al., EMNLP 2022), the citation-based reference.
You: generate 2–3 summaries per abstract for ~10k testbed documents with a local 3B model, fine-tune BGE-small, evaluate on the scientific-article set and SciFact, and plot alignment/uniformity against a SimCSE-dropout baseline.

8. PaECTER: Patent-Level Representation Learning Using Citation-Informed Transformers (Ghosh et al., 2024) · recent · ★★ · Colab · training

https://arxiv.org/abs/2402.19411 · model and data: huggingface.co/mpi-inno-comp/paecter
SPECTER for patents: BERT-for-Patents fine-tuned with examiner citations as positives. The open reference encoder for prior-art search; the 2026 Benchmarking Patent Embeddings study (arXiv:2605.24297) finds that "different sections of the same patent as positives" gives the largest gain of all recipes.
You: fine-tune BGE-small on a 20–50k subset of the PaECTER triplets and show the gain on held-out citation retrieval; then apply the same recipe with citations as positives on the scientific-article testbed and compare with SPECTER2.

9. DAPFAM and PatenTEB: a patent retrieval benchmark and a patent embedding benchmark (Ayaou et al., 2025) · recent · ★★ · laptop/Colab

https://arxiv.org/abs/2506.22141 · https://arxiv.org/abs/2510.22264 · code: github.com/iliass-y/patenteb
BEIR and MTEB for patents: family-level prior-art retrieval with in-domain vs out-of-domain partitions (out-of-domain recall about 5× lower), passage beats document level, dense only modestly ahead of BM25, RRF a strong cheap fusion.
You: run the spring pipeline on DAPFAM in-domain vs out-of-domain and confirm the gap and the RRF gain; then test whether patent-trained patembed models help or hurt on SciFact/NFCorpus.

10. GPL: Generative Pseudo Labeling for Unsupervised Domain Adaptation of Dense Retrieval (Wang et al., NAACL 2022) · classic · ★ · Colab · training

https://arxiv.org/abs/2112.07577 · code: github.com/ukplab/gpl
doc2query generates queries for the target corpus, a cross-encoder scores them, the bi-encoder learns to reproduce the margins. Up to +9 nDCG@10 on specialised BEIR sets. Extension: R-GPL (2025) re-mines negatives as the model adapts; NV-Retriever (2024) filters false negatives.
You: run GPL on SciFact/NFCorpus with the spring generator and cross-encoder as teachers and BGE-small as student; report zero-shot vs GenQ vs GPL, then add negative re-mining and positive-aware filtering.

11. BGE-M3: Multi-Linguality, Multi-Functionality, Multi-Granularity Text Embeddings (Chen et al., Findings of ACL 2024) · recent · ★★ · Colab

https://arxiv.org/abs/2402.03216 · model: BAAI/bge-m3
One 100-language encoder that outputs dense, sparse and ColBERT-style vectors at once, trained by distilling the fused score into each head. The open multilingual baseline.
You: reproduce the dense/sparse/multi-vector ablation and fusion on the testbed, then translate SciFact queries to French and measure what each head loses cross-lingually; compare with the spring RRF pipeline.

12. Timeline talk: Unsupervised word embeddings capture latent knowledge from materials science literature (Tshitoyan et al., Nature 2019) · classic · ★★ · laptop

https://www.nature.com/articles/s41586-019-1335-8 · code: github.com/materialsintelligence/mat2vec
word2vec trained on abstracts up to year T ranks compounds by similarity to "thermoelectric"; many top candidates are only reported as thermoelectrics years later. Embeddings as a way to watch a field evolve, and possibly to anticipate it. Modern relatives: diachronic periodical embeddings (Lyu & Ke 2025), TechToken for patents (2026).
You: redo the backtest on one subfield year by year (arXiv cs.IR or the scientific testbed) with word2vec and with a document encoder, and plot subtopic centroid trajectories over time. Prediction or hindsight?

Session 4 (Nov 27) — LLMs in the loop: rerankers, rewriters, reasoning

13. Is ChatGPT Good at Search? LLMs as Re-Ranking Agents (RankGPT) (Sun et al., EMNLP 2023) · classic · ★ · Colab (+ optional API)

https://arxiv.org/abs/2304.09542 · code: github.com/sunnweiwei/rankgpt
Ask the LLM to output a permutation of passage IDs, sliding-window over the top-100. Beats supervised cross-encoders, and a 440M model distilled from its permutations beats a 3B supervised reranker. The Setwise paper (Zhuang et al., SIGIR 2024) compares pointwise, pairwise, listwise and setwise prompting on cost and quality.
You: rerank BM25 top-50 on SciFact/NFCorpus with an open 7B model in 4-bit (or RankZephyr) against the spring cross-encoder; measure the four prompting paradigms' cost/quality front with Flan-T5; then distil permutations into MiniLM.

14. Rank1: Test-Time Compute for Reranking (Weller et al., COLM 2025) · recent · ★★ · Colab

https://arxiv.org/abs/2502.18418 · code: github.com/orionw/rank1 (0.5B–32B models)
A reranker that "thinks": reasoning traces from DeepSeek-R1 distilled into Qwen models that write a chain of thought, then a relevance token. State of the art on reasoning-heavy retrieval; explanations come for free. Counter-paper: Don't Overthink Passage Reranking (2025) finds the thinking sometimes hurts.
You: run Rank1-0.5B/1.5B/3B on the testbed against the MiniLM cross-encoder, count generated tokens per pair, then disable the reasoning block and compare. Is the explanation worth the cost?

15. HyDE: Precise Zero-Shot Dense Retrieval without Relevance Labels (Gao et al., ACL 2023) · classic · ★ · laptop + small LLM

https://arxiv.org/abs/2212.10496 · code: github.com/texttron/hyde · critique: When do generative query and document expansions fail? (Weller et al., EACL 2024, https://arxiv.org/abs/2309.08541)
Let an LLM write a hypothetical answer, embed that instead of the query. The critique shows expansions help weak retrievers and hurt strong ones, and mostly help under distribution shift.
You: reproduce HyDE and Query2doc on the testbed with BGE-small and a local 3B generator, then reproduce the "weak vs strong ranker" interaction across BM25, MiniLM, BGE and the cross-encoder. Does a hypothetical abstract beat a hypothetical answer on scientific queries?

16. BRIGHT: A Realistic and Challenging Benchmark for Reasoning-Intensive Retrieval (Su et al., ICLR 2025) · recent · ★★ · Colab

https://arxiv.org/abs/2407.12883 · code: github.com/xlang-ai/BRIGHT · leaderboard brightbenchmark.github.io
1,384 real questions (StackExchange, LeetCode, maths) where relevance needs reasoning, not overlap. The best MTEB retriever drops to 18 nDCG@10; LLM query rewriting adds 12 points; the leaderboard went from 18 to 67 in two years.
You: reproduce the baseline table on three small subsets with BM25, BGE-small and the spring pipeline; reproduce the reason-then-retrieve gain with a 7B rewriter; test whether reasoning rewrites help on ArguAna and the scientific-article set.

17. Transformer Memory as a Differentiable Search Index (DSI) (Tay et al., NeurIPS 2022) · classic · ★★ · Colab · training

https://arxiv.org/abs/2202.06991 · code: github.com/ArvinZhuang/DSI-QG (DSI trainer) · scaling study: How Does Generative Retrieval Scale to Millions of Passages? (Pradeep et al., EMNLP 2023, https://arxiv.org/abs/2305.11841)
One T5 memorises the corpus: document text → docid at indexing time, query → docid at query time, beam search over docids. The concrete form of "the model is the index". The scaling study finds synthetic queries are the one essential ingredient and that the approach falls far behind dual encoders at 8.8M passages.
You: reproduce DSI on NQ10K with T5-small (semantic vs string docids), then build the scaling curve on the MS MARCO subset (1k / 10k / 92k passages) with doc2query pseudo-queries, against BM25 and MiniLM. What is generative retrieval adding if it needs doc2query to work?

18. From RankNet to LambdaRank to LambdaMART (Burges, Microsoft Research 2010) · classic · ★ · laptop

https://www.microsoft.com/en-us/research/publication/from-ranknet-to-lambdarank-to-lambdamart-an-overview/ · code: LightGBM lambdarank objective · companion: Are Neural Rankers still Outperformed by Gradient Boosted Decision Trees? (Qin et al., ICLR 2021, https://openreview.net/forum?id=Ut1vF_q_vC)
Learning to rank with gradient-boosted trees: pairwise gradients rescaled by the nDCG change of swapping two documents. Won the Yahoo! LTR challenge in 2010 and still runs most production search engines, including patent search. The 2021 audit finds most neural rankers lose to it on feature datasets.
You: reproduce LambdaMART on MSLR-WEB10K with LightGBM (pointwise vs lambdarank), then train a feature-based ranker on the MS MARCO subset with BM25, dense, cross-encoder and RRF scores as features. Does learned fusion beat RRF, and which feature matters?

Session 5 (Dec 4) — RAG and agents

19. Self-RAG: Learning to Retrieve, Generate, and Critique through Self-Reflection (Asai et al., ICLR 2024) · classic · ★★★ · Colab

https://arxiv.org/abs/2310.11511 · code: github.com/AkariAsai/self-rag (7B/13B models)
One LM emits reflection tokens (retrieve?, is the passage relevant?, is the answer supported?) and decodes with them as a segment-level score. The canonical adaptive-RAG paper. Lighter alternative for the same slot: CRAG (Yan et al., 2024), a T5 evaluator that triggers keep/discard/web-fallback.
You: run selfrag_llama2_7b in 4-bit on a PopQA subset with retrieval on/off; then use its relevance and support token probabilities as a reranker on SciFact/NFCorpus against the spring cross-encoder.

20. RAPTOR: Recursive Abstractive Processing for Tree-Organized Retrieval (Sarthi et al., ICLR 2024) · recent · ★★ · Colab

https://arxiv.org/abs/2401.18059 · code: github.com/parthsarthi03/raptor · rebuttal: Stronger Baselines for RAG with Long-Context LMs (Laitenberger, Manning & Liu, EMNLP 2025, https://arxiv.org/abs/2506.03989)
Cluster chunks, summarise each cluster with an LLM, recurse; search leaves and summaries together. +20 points on QuALITY with GPT-4. The rebuttal says a simple order-preserving retrieve-then-read matches it.
You: reproduce collapsed-tree vs leaf-only retrieval on a QuALITY subset with MiniLM embeddings and a 1.5B summariser; build a tree over NFCorpus and report nDCG@10 for leaves, tree and the DOS-RAG baseline, plus indexing cost.

21. Dense X Retrieval: What Retrieval Granularity Should We Use? (Chen et al., EMNLP 2024) · recent · ★ · laptop/Colab

https://arxiv.org/abs/2312.06648 · code: github.com/chentong0/factoid-wiki
Index propositions (atomic self-contained facts) instead of passages or sentences; recall@5 improves for every retriever tested. The chunking cluster: Anthropic's Contextual Retrieval, late chunking, and Is Semantic Chunking Worth the Computational Cost? (Qu et al., 2024).
You: propositionise SciFact/NFCorpus with the released model; compare passage, sentence, proposition, LLM-written contextual headers and late chunking with BGE-small. Who pays for a 6× larger index?

22. The Power of Noise: Redefining Retrieval for RAG Systems (Cuconasu et al., SIGIR 2024) · recent · ★ · laptop/Colab

https://arxiv.org/abs/2401.14887 · code: github.com/florin-git/The-Power-of-Noise · reproduction: The Powerless Noise (Mazuryk et al., SIGIR 2026 reproducibility track, https://arxiv.org/abs/2607.03615)
Related-but-wrong distractors hurt the reader; random documents surprisingly help. The 2026 reproduction finds the effect appears, weakens or vanishes with small changes of model, prompt and output constraints. Pair with Lost in the Middle (Liu et al., TACL 2024) for position effects.
You: reproduce the random-document gain with Phi-2 or Qwen2.5-1.5B on the repo's NQ subset, then apply the 2026 paper's perturbations; on SciFact, reader accuracy with gold + k random vs gold + k hard negatives from the BM25/dense runs.

23. Search-R1: Training LLMs to Reason and Leverage Search Engines with Reinforcement Learning (Jin et al., 2025) · recent · ★★★ · Colab (inference)

https://arxiv.org/abs/2503.09516 · code: github.com/PeterGriffinJin/Search-R1 (3B/7B checkpoints)
RL teaches a 3B/7B model to interleave thinking and search calls against a local index, with only an exact-match reward. +20–40 % over RAG on seven QA sets. Training-free alternative: Search-o1 (Li et al., EMNLP 2025). Classic ancestors: ReAct (2023), IRCoT (2023).
You: no training: run Search-R1-3B with a local BM25/FAISS index on HotpotQA and SciFact, compare with a ReAct/IRCoT loop on Qwen3-4B, count searches per question and recall of the union of issued queries. Is the gain reasoning, or query expansion by another name?

Session 6 (Dec 11) — Trust: evaluation, attacks, bias, and documents that are not text

24. Can LLMs replace relevance assessors? Thomas et al. (SIGIR 2024) · UMBRELA (2024) · Clarke & Dietz (2024/26) · recent · ★★ · laptop/Colab

https://arxiv.org/abs/2309.10621 · https://arxiv.org/abs/2406.06519 · https://arxiv.org/abs/2412.17156 · code: github.com/castorini/umbrela
Bing finds GPT-4 labels as accurate as humans and cheaper; UMBRELA reproduces this and became the TREC RAG judge; Clarke & Dietz answer that LLM judges cap what can be measured and favour LLM rerankers ("the inherent narcissism of LLMs").
You: judge the testbed top-20 with Qwen2.5 1.5B/3B/7B using the UMBRELA prompt; report agreement with human qrels and the Kendall τ of system rankings; then reproduce the circularity argument by judging an LLM-reranked run with the same model. Where does a small judge stop being useful?

25. eRAG: Evaluating Retrieval Quality in Retrieval-Augmented Generation (Salemi & Zamani, SIGIR 2024) · recent · ★★ · Colab

https://arxiv.org/abs/2404.13781 · code: github.com/alirezasalemi7/eRAG
Human relevance labels correlate poorly with what the generator actually needs. Feed each retrieved document alone to the LLM, score the output, use that as the document's relevance; τ with end-to-end RAG performance improves by up to 0.49. Related: RAGAS, RAGChecker, TREC nugget evaluation.
You: on the MS MARCO subset, compute human-qrel nDCG@10, eRAG nDCG@10 and end-to-end answer quality for BM25, MiniLM and the hybrid pipeline; report the three pairwise τ. Is an IR test collection still reusable if relevance is generator-specific?

26. PoisonedRAG: Knowledge Corruption Attacks to RAG of LLMs (Zou et al., USENIX Security 2025) · recent · ★★ · Colab

https://arxiv.org/abs/2402.07867 · code: github.com/sleeepeer/PoisonedRAG · defence: RobustRAG (Xiang et al., 2024, github.com/inspire-group/RobustRAG)
Five crafted texts per target question make a RAG system answer as the attacker wants, ~90 % of the time, even black-box. Classic ancestor: Poisoning Retrieval Corpora by Injecting Adversarial Passages (Zhong et al., EMNLP 2023). Local corpora and local models only.
You: reproduce the black-box attack on the local MS MARCO subset with a local 3–7B generator; vary retriever (BM25, dense, RRF, +reranker) and k; then apply RobustRAG's isolate-then-aggregate defence and report clean vs certified accuracy. Why does hybrid fusion act as a defence?

27. Neural Retrievers Are Biased Towards LLM-Generated Content (Dai et al., KDD 2024) · recent · ★ · laptop + small LLM

https://arxiv.org/abs/2310.20501 · code: github.com/KID-22/Source-Bias · explanation: Perplexity Trap (Wang et al., ICLR 2025, https://arxiv.org/abs/2503.08684)
Rewrite a corpus with an LLM, keep the meaning: dense retrievers and rerankers rank the LLM version higher, BM25 does not. The 2025 paper shows retrievers learned perplexity as a relevance feature and proposes a causal correction. What happens to scientific search as abstracts get LLM-polished?
You: paraphrase 500 relevant NFCorpus/SciFact documents with a local 3B model, index both versions, measure the rank shift for BM25, MiniLM, BGE-small and the cross-encoder; then check whether a small LLM judge also prefers the rewrites, and run a 3-round "spiral of silence" loop.

28. ColPali: Efficient Document Retrieval with Vision Language Models (Faysse et al., ICLR 2025) · recent · ★★ · Colab · own PDF testbed

https://arxiv.org/abs/2407.01449 · code: github.com/illuin-tech/colpali · models huggingface.co/vidore (ColQwen2, 2B, fits a T4)
Feed page images to a VLM, keep all patch embeddings, score with ColBERT-style MaxSim. No OCR, no parsing; beats the best text pipeline by 14 points on ViDoRe. Companion: OCR Hinders RAG (Zhang et al., ICCV 2025) quantifies what parsing costs.
You: index ~300 pages of scientific PDFs with ColQwen2 (10–25 min on a T4), write 50 page-level questions, and compare MaxSim vs a mean-pooled single vector vs BM25 over Docling text; measure index size and latency. Is the gain from vision or from the 2B backbone?

Alternates (ask if you would rather present one of these)

MUVERA: multi-vector retrieval via fixed-dimensional encodings (NeurIPS 2024, a numpy project) · RaBitQ 1-bit quantisation with error bounds (SIGMOD 2024) · answerai-colbert-small: train your own late-interaction model by distillation (2024/25) · LitSearch: a literature-search benchmark (EMNLP 2024) · OpenScholar: RAG over 45M papers (Nature 2026) · NV-Retriever: what is a good hard negative? (2024) · Embedding compression improves domain adaptation (2026, laptop) · Search-o1 / BrowseComp-Plus / DCI-Agent (grep-based retrieval for agents, 2026) · HippoRAG 2 (ICML 2025) · RAG vs long context and Self-Route (EMNLP 2024) · Vec2Text: embeddings reveal the text (EMNLP 2023) · Fairness of exposure in rankings (KDD 2018) · Indirect prompt injection (Greshake et al., 2023) · TempRetriever: time-aware dense retrieval (WSDM 2026) · HistWords: diachronic word embeddings (ACL 2016) · OCR Hinders RAG (ICCV 2025) · CoIR: code retrieval benchmark (2024/25) · TAS-B: distil a cross-encoder into a bi-encoder with Margin-MSE (SIGIR 2021) and Beyond Hard Negatives (SIGIR 2026) · Unbiased LTR meets reality: the Baidu click dataset (SIGIR 2024) · Generative retrieval as multi-vector dense retrieval (SIGIR 2024) · TIGER: semantic IDs for generative recommendation (NeurIPS 2023) · LongEmbed: extending embedding models to long context (Zhu et al., EMNLP 2024, https://arxiv.org/abs/2404.12096) · SEAL: substrings as document identifiers (Bevilacqua et al., NeurIPS 2022, https://arxiv.org/abs/2204.10628) · Found in the Middle: permutation self-consistency against position bias in LLM rankers (Tang et al., NAACL 2024, https://arxiv.org/abs/2310.07712).