Modern coding agents are usually evaluated by whether they eventually produce a correct patch, but patch generation depends on an earlier context-acquisition stage: finding the repository files needed for the task. We introduce Agent Retrieval Bench, a file-level benchmark for this upstream retrieval problem. Samples are built from real coding-workflow signals and evaluated against frozen base-commit repositories, with relevance defined by what an agent needs next rather than direct query-file semantic similarity. The benchmark covers four positive-retrieval tasks: code2test, comment2context, trace2code, and edit2ripple; a fifth subset evaluates selective retrieval using natural evidence-backed no-gold cases and counterfactual wrong-repository controls. Agent Retrieval Bench contains 427 samples across 25 repositories: 345 positive examples, 50 natural no-gold examples, and 32 counterfactual controls. The corpus includes 308 base-commit snapshots, 392,000 files, and 7.9 million chunks. We evaluate lexical retrieval, RepoMap, open-source embeddings, selective abstention, and logged agent context selection. No single retrieval family dominates: Qwen3-Embedding-4B has the best sample-weighted MRR on positive samples, Qwen3-Embedding-8B the best Recall@20, and RepoMap the best budgeted context yield at 8K tokens, with task-level winners differing substantially. Selective thresholds calibrated with counterfactual controls do not improve selective success on natural no-gold cases, revealing a calibration gap. Logged trajectories also miss every gold file on 27-35 percent of samples. A controlled seed-intervention pilot finds that retrieval-derived initial context yields higher file F1 with less post-seed exploration than random non-gold context, while oracle gold context shows substantial remaining headroom.
Retrieval-augmented generation (RAG) typically treats context selection as ranking chunks against a single query embedding. This assumption breaks down for complex queries, such as multi-hop or ambiguous questions, where top-k selection tends to over-cover one semantic aspect while ignoring critical sub-questions. We propose GeoRAG, which recasts context selection as Information Demand Coverage Optimization. GeoRAG builds a multi-dimensional demand distribution through diverse sub-query generation and reverse-validation weighting, then selects context by minimizing the Sinkhorn-Wasserstein distance between this demand distribution and the coverage of the selected set. The resulting demand-weighted facility-location objective is monotone submodular, giving a $1-1/e$ greedy guarantee, which we approximate with a Sinkhorn-based marginal-gain surrogate. The method is unsupervised, training-free, and retrieval-agnostic. We further show that single-point, query-proximity scorers cannot cover multi-modal demands, exposing a structural limit of ranking-based selection. On six open-domain QA benchmarks, GeoRAG improves exact match (EM) by +6.5 to +7.5 points over top-k truncation (up to +9.7 on HotpotQA and ASQA) and outperforms strong baselines including MMR, DPP, BGE-Reranker, SMART-RAG, and AdaGReS, with stable gains across context budgets and sub-query generators.
A modern coding agent can hold an entire repository in its context window. Most of its reading is wasted -- and the interesting question is not how much context an agent can use, but what it actually \emph{needs}. We study that question at the moment it matters most: when the agent must \emph{edit} code. Separating \emph{finding} the work site from \emph{acting} on it, we hold localization fixed with an oracle, vary only how the code is represented, and score context against real issue resolution on SWE-bench Verified. The answer is starkly minimal. The signal lives in the code being edited itself: natural-language summaries of it answer almost none of the behavioral questions that the source answers ($4/45$ vs.\ $27/45$, held-out repositories, independent judge), and the gap belongs to the representation, not the summarizer -- a frontier model's summaries score exactly as poorly as a 3B model's. The surrounding context hardly matters either: across every multi-file instance in Verified, under a protocol frozen before any data, rendering a file's remainder as UML skeletons and signatures resolves no more issues than deleting that remainder outright ($N{=}70$, exact McNemar $p{=}0.75$). That was our registered hypothesis, and it failed. Compressed context, meanwhile, matches whole files at a third of the tokens: a resolved issue costs $19$K context tokens, not $94$K. The instrument also yielded a finding the field should keep: temperature-0 API inference flips ${\sim}9\%$ of per-instance outcomes between byte-identical runs. That is a noise floor under every small effect reported on this benchmark, including ours. We release the instrument -- gold-validated environments, per-instance proof that every reference edit is expressible from every arm's context, deterministic patch construction, and pre-registered hypotheses whose nulls we publish.
Electronic health records now routinely exceed 100,000 tokens per patient. Yet large language models exhibit the lost-in-the-middle (LitM) effect: information near the center of a long context is retrieved less reliably than information near the edges. In clinical use this is not benign: the single most consequential fact in a note can sit at its center. We term this the clinical lost-in-the-middle (CLitM) problem, give its first systematic characterization using MedAlign, and compare context-selection strategies as remedies. Across 2,196 instruction-response pairs and six language models, we observe a 21.9 percentage-point gap between peak accuracy (59.5%, 95% CI [46.3, 71.0], 20-30% decile) and trough accuracy (37.6% [23.2, 52.5] at 70-80%); 67.8% of reference answers fall between the 10th and 90th percentiles of the EHR timeline, inside the CLitM trough. We introduce Query-Conditioned Clinical Suppression (QCCS), a lightweight query-conditioned selection gate, and evaluate it against BM25, BM25 with section-header filtering, dense retrieval, and cross-encoder reranking (N=83 held-out instructions). With Qwen2.5-7B-Instruct (16k context), QCCS outperforms all five comparators under LLM-as-judge scoring: for middle-position instructions QCCS reaches 16.7% versus BM25 3.3%, cross-encoder 0.0%, dense 0.0%, and full context 6.7%; overall QCCS reaches 25.3% versus at most 3.6% for retrieval-only comparators. This advantage is not explained by retrieval recall: at k=20, BM25 retrieves the gold evidence sentence in 98.8% of instructions (QCCS 34.9%), yet retrieval arms stay at most 2.6% accurate even when they retrieve it, whereas QCCS reaches 25.0% even when it does not. In this proof-of-concept evaluation, query-aligned context selection predicts EHR instruction-following accuracy better than gold-sentence retrieval recall.
Jamie Heredge, Mattia J. Villani, Pranav Deshpande +2cs.LG cs.AI stat.ML
Prior-fitted networks (PFNs) are a promising class of tabular foundation models that perform in-context learning, whereby the entire labelled training set is supplied as context, and predictions for test queries are produced in a single forward pass. However, the quadratically scaling self-attention mechanism in many PFN architectures makes inference prohibitive for very large training datasets. We propose CRUMB (Clustered Retrieval Using Minimised-MMD Batching), a three-stage inference wrapper that (i) clusters the test queries, (ii) selects a small, distributionally matched training subset for each cluster by greedily minimising the maximum mean discrepancy (MMD), and (iii) runs exact PFN inference on each reduced-context batch. CRUMB is architecture-agnostic and requires no retraining. On the 51-dataset TabArena benchmark, evaluated across three PFN architectures (TabPFNv2, TabICLv1, TabICLv2), we show that CRUMB outperforms similar state-of-the-art context selection strategies. We also show that CRUMB is resilient to covariate drift, as the MMD-minimisation step naturally helps align the training context distribution to match the current test batch distributions.
Modern large language model (LLM) agents do not simply need longer contexts; they need decision-relevant evidence at the moment of action. We study decision-aware context selection: ranking retrieved files, tests, traces, rules, and memories by their expected effect on an agent's next action rather than by semantic similarity alone. We present the Counterfactual-Inspired Context Layer (CICL), which builds an instance context graph, estimates decision-oriented utility for candidate units, and compresses selected evidence into typed memory cards. The same schema can be instantiated with hosted LLM judges, local surrogates, or lightweight rankers, making the selection protocol auditable across model choices. On 50 SWE-bench Verified file-retrieval instances, Qwen3.6-Plus reranking of BM25 top-50 candidates improves hit@1 from 0.58 to 0.78 and MRR@10 from 0.634 to 0.790, with all 2,500 judgments parseable. Controlled diagnostics show that CICL identifies action-critical evidence: removing the top-utility semantic unit reduces F1 from 0.245 to 0.000. In selected-then-compressed mode, memory cards save 44.93 tokens per query while preserving selected evidence. CICL provides a practical layer for measuring, ranking, and compressing decision-critical context for tool-using agents. Code is available at https://github.com/stephen-guan-researcher/CICL.
Effective document-level machine translation (DocMT) requires capturing long-range discourse dependencies. Recent work has explored retrieval-based and discourse-aware context selection. However, these approaches often lack an explicit mechanism for modeling structured discourse dependencies between distant paragraphs in a document. In this paper, we propose G^2C-MT (Graph-Guided Context for Machine Translation), which views DocMT context selection as a structured path discovery problem on a lightweight discourse graph, rather than retrieving unstructured context sets or relying on expensive LLM-based discourse modeling. In detail, we represent each paragraph as a node and model the relationship between each pair of nodes, considering their semantic similarity, adjacency, and keyword overlap. Furthermore, we propose a depth-biased random walk over the graph to sample a backward context path for each target paragraph. The context path will be used to prompt a large language model (LLM) for translation. This framework naturally supports multi-path context sampling, which can improve robustness by aggregating diverse translation candidates for discourse-ambiguous inputs. Experiments conducted across various domains show that G^2C-MT outperforms strong baselines on multiple LLMs, including DeepSeek-V3, Gemini-2.5-Flash-lite, and the Qwen-2.5/3 series.