PaperlayerView exact revision on arXiv

Source-linked AI summary

Beyond Semantic Similarity: Rethinking Retrieval for Agentic Search via Direct Corpus Interaction

Zhuofeng Li, Haoxiang Zhang, Cong Wei, Pan Lu, Ping Nie, Yi Lu, Yuyang Bai, Shangbin Feng, Hangxiao Zhu, Ming Zhong, Yuyu Zhang, Jianwen Xie, Yejin Choi, James Zou, Jiawei Han, Wenhu Chen, Jimmy Lin, Dongfu Jiang, Yu Zhang

arXiv:2605.05242v1cs.IRcs.AI

TL;DR

Conventional top-k retrieval interfaces limit agents’ access to corpus evidence during multi-step search. This paper studies direct corpus interaction, in which agents search raw corpora with terminal tools, and finds strong performance across agentic search, multi-hop QA, and ranking benchmarks without conventional retrievers.

  • Problem

    Fixed top-k retrieval interfaces constrain agents’ ability to combine sparse clues, enforce exact lexical constraints, inspect local context, and revise multi-step search plans.

  • Method

    Direct corpus interaction lets agents search and verify raw corpora with general-purpose terminal tools without embedding models, vector indexes, or retrieval APIs.

  • Results

    80.0% accuracy on BrowseComp-Plus, up from 69.0% with the matched conventional retriever, while reducing cost by 29.4%; DCI also outperformed baselines across multi-hop QA and ranking.

  • Takeaways & Limitations

    Trajectory analyses indicate that DCI’s advantage comes from higher-resolution local inspection, verification, and compositional search of surfaced evidence.

  • Takeaways & Limitations

    DCI’s accuracy drops as corpus breadth grows, with higher tool usage, latency, and cost, and severe degradation at 400K documents.

Abstract

from arXiv · show

Modern retrieval systems, whether lexical or semantic, expose a corpus through a fixed similarity interface that compresses access into a single top-k retrieval step before reasoning. This abstraction is efficient, but for agentic search, it becomes a bottleneck: exact lexical constraints, sparse clue conjunctions, local context checks, and multi-step hypothesis refinement are difficult to implement by calling a conventional off-the-shelf retriever, and evidence filtered out early cannot be recovered by stronger downstream reasoning. Agentic tasks further exacerbate this limitation because they require agents to orchestrate multiple steps, including discovering intermediate entities, combining weak clues, and revising the plan after observing partial evidence. To tackle the limitation, we study direct corpus interaction (DCI), where an agent searches the raw corpus directly with general-purpose terminal tools (e.g., grep, file reads, shell commands, lightweight scripts), without any embedding model, vector index, or retrieval API. This approach requires no offline indexing and adapts naturally to evolving local corpora. Across IR benchmarks and end-to-end agentic search tasks, this simple setup substantially outperforms strong sparse, dense, and reranking baselines on several BRIGHT and BEIR datasets, and attains strong accuracy on BrowseComp-Plus and multi-hop QA without relying on any conventional semantic retriever. Our results indicate that as language agents become stronger, retrieval quality depends not only on reasoning ability but also on the resolution of the interface through which the model interacts with the corpus, with which DCI opens a broader interface-design space for agentic search.

1 Introduction

The introduction argues that fixed top-k retrieval interfaces constrain agentic search by filtering evidence before reasoning, and proposes direct corpus interaction (DCI) as a higher-resolution alternative. DCI searches raw corpora with general-purpose terminal tools and achieves strong results across agentic search, multi-hop QA, and ranking-oriented IR settings.

  • Introduction: Conventional retrieval constrains iterative retrieval agents to a fixed interface exposing only a top-k corpus slice, making exact constraints, sparse clue combinations, and multi-step evidence refinement difficult.The limitation is especially pronounced for emerging agentic-search benchmarks.
  • Introduction: DCI gives agents direct access to the raw corpus through grep, file reads, shell commands, and lightweight scripts, without embedding models, vector indexes, retrieval APIs, or top-k mediation.This leaves the entire corpus available for agent-directed search and local evidence inspection.
  • Introduction: Replacing Qwen3-Embedding-8B retrieval with DCI under the same Claude Sonnet 4.6 backbone raises BrowseComp-Plus accuracy from 69.0% to 80.0% (+11.0 points) while reducing cost from $1,440 to $1,016 (−29.4%).The evaluation covers BrowseComp-Plus, multi-hop QA, and ranking-oriented IR benchmarks.
  • Introduction: Trajectory and ablation analyses indicate that DCI’s gains do not primarily arise from surfacing more gold documents, and that its advantage persists under a highly restricted tool profile.The studies also examine evidence use, corpus scale, context management, and tool usage to characterize DCI’s operating envelope.
  • Introduction: The paper formalizes DCI as a retrieval paradigm, evaluates it across diverse agentic settings, and introduces retrieval interface resolution as a lens for explaining its effectiveness.The lens emphasizes access to units smaller and more precise than entire documents or passages.

2 Related Work

Related work progresses from classical retrieve-then-generate RAG toward multi-step retrieval agents that iteratively search, accumulate evidence, and refine plans. Coding-agent research motivates direct interaction with raw corpora through command-line tools, while this work frames it as a broader retrieval interface for agentic search.

  • RAG: Classical RAG retrieves potentially relevant documents from a corpus before an LLM conditions on the returned evidence to generate an answer (Lewis et al., 2020; Borgeaud et al., 2022; Ram et al., 2023; Gao et al., 2023; Shi et al., 2024).
  • The Rise of Agentic Search: Recent agentic-search systems replace single-shot retrieval with multi-step investigations that iteratively search, accumulate evidence, and refine plans from intermediate observations (Jin et al., 2025; Jiang et al., 2025; Li et al., 2025).
  • Coding Agents: Command-line coding agents demonstrate that bash, file search, code editing, grep, and reading tools can substantially improve complex software-engineering and code-localization tasks (Liu et al., 2023; Jimenez et al., 2024; Deng et al., 2025; Anthropic, 2025; Merrill et al., 2026; Li et al., 2026a; Yang et al., 2024; Sutawika et al., 2026).
  • Direct Corpus Interaction: Direct corpus interaction bypasses offline preprocessing and a separate retriever, letting agents search raw corpora with terminal tools and lightweight scripts rather than reasoning only over returned top-k candidates.
  • Coding Agents: Prior document-QA work finds that a tool-augmented keyword-search agent over raw PDFs can approach vector-database RAG, whereas this paper studies direct corpus interaction as a broader retrieval interface for agentic search (Subramanian et al., 2025).

3 Direct Corpus Interaction

Direct corpus interaction (DCI) bypasses conventional retrieval interfaces, letting agents search raw corpora through terminal tools and file reads. The section also defines runtime scaffolds and trajectory metrics for evaluating evidence access and localization.

  • Direct Corpus Interaction: DCI bypasses embedding models, vector indexes, and retrieval APIs, exposing raw corpora through grep, rg, find, glob, file reads, and lightweight scripts.This produces tool outputs directly, including exact or regular-expression matches and local context around matches.
  • Agent Scaffolds: DCI-Agent-Lite and DCI-Agent-CC implement the same raw-corpus interface under minimal and stronger runtime scaffolds, separating interface effects from harness engineering.The stronger scaffold adds prompting, tool orchestration, and context handling, while the minimal scaffold contains no retrieval-specific module, offline index, dense retriever, or reranker.
  • Runtime Context Management: Runtime context management preserves long-horizon search structure through truncation, zero-LLM compaction, and model-generated summarization under increasing context pressure.These policies alter how much tool-mediated evidence remains in working context, not the retrieval interface itself.
  • Evaluation Metrics: Coverage measures whether trajectories surface gold evidence, ranging from reaching any gold document to surfacing the complete gold set.coverageany tests whether at least one gold document is surfaced, coveragemean averages surfaced gold documents, and coverageall tests whether all are surfaced.
  • Evaluation Metrics: Localization measures whether, after reaching a useful gold document, the agent narrows to a small, concentrated evidence span within it.It is a within-document metric based on the best segment-level score for each surfaced gold document, with higher scores indicating more concentrated evidence extraction.

4 Experiments

Across agentic search, knowledge-intensive QA, and IR ranking, direct corpus interaction substantially outperforms conventional retrieval interfaces while remaining cost-efficient on BrowseComp-Plus. The experiments attribute these gains to fine-grained evidence discovery and use, while identifying rising costs with broader corpora and sensitivity to agent design.

  • RQ1: DCI-Agent-CC reaches 80.0% on BrowseComp-Plus, 83.0% average QA, and 68.5% average NDCG@10, surpassing retrieval baselines across tasks.It exceeds ASearcher-Local-14B by 30.7 points in QA and ReasonRank-32B by 21.5 points in IR ranking; Table 2 and Table 3 report these comparisons.
  • RQ1: With the same Claude Sonnet 4.6 backbone, DCI raises BrowseComp-Plus accuracy from 69.0% to 80.0% while reducing cost from $1,440 to $1,016.DCI-Agent-CC also surpasses GPT-5 + Qwen3-Embedding-8B at 71.7%.
  • RQ2: DCI’s advantage comes from fine-grained discovery, composition, and evidence use rather than primarily higher gold-document recall.DCI-Agent-CC correctly answers 176 questions missed by the matched retrieval agent, while the reverse occurs for 76; 142 missed cases already surface at least one gold document.
  • RQ3: DCI-Agent-Lite trades lower mean gold-document coverage for much higher localization, with scores of 28.0 versus 56.7 and 48.4 versus 21.7.Its coverageany score remains comparable to Qwen3-Embedding-8B’s, at 70.0 versus 74.0.
  • RQ4: Expanding the corpus from 100K to 200K documents increases DCI-Agent-CC’s tool calls from 38.5 to 86.9, more than doubles latency and cost, and lowers accuracy.These results define an operating envelope in which DCI scales better with search depth than search breadth.
  • RQ5–RQ6: Runtime context management is non-monotonic, while a constrained read-plus-grep interface already achieves 61% BrowseComp-Plus accuracy.More aggressive context management does not simply improve results, indicating that both policy and tool expressivity materially affect DCI-Agent-Lite.

5 Conclusion … Sparse & Dense Retrieval

The paper presents direct corpus interaction (DCI) as a retrieval paradigm in which agents search and verify raw corpora with terminal tools rather than conventional retrievers. The supplied appendix passages define the evaluated benchmarks, baselines, proprietary backbones, retrieval agents, and sparse, dense, and reranking systems.

  • 5 Conclusion: DCI enables agentic search against raw corpora with general-purpose terminal tools, without conventional retrievers or offline indexes, and is competitive across ranking, multi-hop QA, and end-to-end agentic search.Controlled ablations and trajectory analyses indicate that DCI’s advantage stems from retri…
  • Appendix; A Experimental Details; A.1 Benchmarks: The evaluation uses full test sets for BRIGHT and Bamboogle, while other datasets use random samples of 50 examples per dataset for efficiency.
  • Agentic Search: BrowseComp-Plus is a closed-corpus benchmark for controlled deep-research-agent evaluation, with human-verified supporting documents, mined hard negatives, and multi-document evidence synthesis.
  • Knowledge-Intensive QA: NQ and TriviaQA evaluate single-hop factual retrieval, while Bamboogle is a manually constructed multi-hop dataset; ambiguous and time-sensitive questions are excluded.
  • IR Ranking: BRIGHT tests reasoning-intensive retrieval across Biology, Earth Science, Economics, and Robotics, while BEIR provides heterogeneous information-retrieval tasks including ArguAna.
  • A.2 Compared Baselines; Proprietary Models: BrowseComp-Plus baseline evaluations use proprietary OpenAI, Claude, GLM, and Kimi models as retrieval-agent backbones, with o3 strongest among baseline agents and GPT-5.4 nano powering DCI-Agent-Lite.
  • Retrieval Agents: Compared retrieval agents include R1-Searcher-7B, Search-R1-32B, ZeroSearch-7B, Verl-Tool-Search-7B-DAPO, and ASearcher-Local-14B, spanning external search, interleaved reasoning, simulated retrieval, tool-use RL, and asynchronous RL.

A.3 Metrics · B Further Discussion of Experimental Results

The evaluation defines accuracy measures for QA and IR, while resolution metrics standardize how retriever outputs and DCI tool traces become observation-level candidates and snippets. DCI-specific reconstruction preserves document coverage but penalizes observations that expose little local evidence, and trajectories rely mainly on compositional search commands.

  • A.3 Metrics: Accuracy is judged with GPT-4.1 for Browsecomp-Plus and knowledge-intensive QA, comparing predicted and reference answers for semantic and numerical correctness.The short, well-defined references make this comparison largely unambiguous.
  • A.3 Metrics: Resolution metrics represent each observation as document–snippet candidates, with snippet lengths determined by the interface’s exposed evidence.These candidates underpin the evaluation of evidence resolution across retriever-mediated access and DCI traces.
  • A.3 Metrics: Retriever-mediated access maps each returned result directly to one candidate, so a top-k call yields n_t = k candidates with preview-length measurements.The retriever API standardizes both candidate sets and preview snippets.
  • A.3 Metrics: DCI evaluation reconstructs candidates by aligning raw tool observations to surfaced gold documents through explicit paths or matched local text.Unlike retriever outputs, DCI traces do not provide candidate sets explicitly and therefore require evaluator-side reconstruction.
  • A.3 Metrics: For grep-style searches, each aligned matched line becomes a candidate, while read-style inspection uses the exposed span when it overlaps annotated gold evidence.Multiple matched lines from one document remain separate candidates, with later scoring retaining the best segment score.
  • A.3 Metrics: Path-only or unmatched observations mark a gold document as surfaced but use the full document as the snippet, yielding low localization scores despite document coverage.This fallback reflects reach without useful evidence localization.
  • A.3 Metrics: Across 100 DCI-Agent-Lite cases, 3,168 commands rely primarily on compositional operators such as rg | head and rg | rg, while full-document reads are rare.The distribution aggregates command usage across both correct and incorrect runs.

B.1 In-Depth Analysis of DCI-Agent-Lite Search Patterns · C Instruction Templates · D Case Study

DCI-Agent-Lite predominantly uses simple, staged shell-search patterns, while the instruction templates enforce corpus-only, iterative, recall-and-precision-aware workflows and structured answer evaluation. A case study presents DCI-Agent-Lite and DCI-Agent-CC as elegant, efficient, and robust deep-research systems.

  • B.1 In-Depth Analysis of DCI-Agent-Lite Search Patterns: Search-plus-limit commands dominate DCI-Agent-Lite tool use at 56.2%, followed by chained searches at 20.6%, with counting, keyword search, listing, scripting, and localization less frequent.Figure 6 characterizes the agent’s Bash interactions by representative command pattern.
  • B.1 In-Depth Analysis of DCI-Agent-Lite Search Patterns: The observed search behavior follows six stages: corpus exploration, broad keyword search, iterative narrowing, and additional later-stage operations described from tool traces.The grouping is ordered roughly chronologically and uses directory scans and ripgrep searches as representative operations.
  • C.2 IR Instruction: The IR template likewise mandates Grep/Bash-only retrieval, parallel and diverse searches, follow-up searches for gaps, and exhaustive exploration rather than stopping after a few documents.It explicitly treats recall and precision as equally important under NDCG evaluation.
  • C.2 IR Instruction: IR instructions require including every genuinely relevant document, carefully excluding irrelevant or tangential documents, and ranking the final list by relevance because ranking quality affects NDCG.The template defines relevance as directly addressing the question or providing essential supporting evidence.
  • C.3 LLM as Judge: The LLM-as-judge template extracts a final answer, compares it only with the provided correct answer, and marks it correct only when it matches or falls within a small numerical margin.It requires focused reasoning about meaningful differences and extracts a confidence score, defaulting to 100% when unavailable.
  • D Case Study: A case study presents DCI-Agent-Lite and DCI-Agent-CC as conducting deep research with greater elegance, efficiency, and robustness.The passage frames the case study around these three qualitative properties.

D.1 Case 1: Demonstration Example

This demonstration shows an agent answering a Natural Questions query through direct Grep and Bash interaction with a closed corpus. It iteratively refines search when scaling limits arise and identifies Don Quixote’s horse as Rocinante from corroborating documents.

  • Agentic reasoning: The agent decomposes a factual query into searchable terms and iteratively refines its retrieval strategy when the initial search encounters scaling limitations.It transitions from broad Grep searches to a targeted Bash command over the corpus.
  • Direct corpus interaction: The agent directly searches the closed wiki corpus with Grep and Bash, enabling precise lexical matching over structured documents without a conventional retrieval agent.The search uses the corpus file directly and retrieves matching content through terminal commands.
  • Answer verification: Rocinante is identified as Don Quixote’s horse, with documents 135082 and 339679 providing direct and corroborating evidence.Document 135082 explicitly describes Rocinante as Don Quixote’s horse, while document 339679 independently confirms the association.

D.2 Case 2a: DCI-Agent-CC on Agentic Search

The DCI-Agent-CC trace solves a twelve-clue BrowseComp-Plus football query by directly inspecting corpus files, verifying match constraints, and refining hypotheses across related documents. It identifies the 1990 England–Belgium match and traces the target entity to AJ Auxerre through interview and player-biography evidence.

  • Case 2a: DCI-Agent-CC on Agentic Search: The agent searches the raw corpus with file listing and reading commands, moving from candidate match files to related Guardian and Wikipedia documents.Its 80-turn trace combines hypothesis formation, constraint verification, and targeted inspection of corpus documents.
  • Case 2a: DCI-Agent-CC on Agentic Search: The agent identifies England’s 1-0 victory over Belgium in the 1990 World Cup Round of 16 as matching the core match constraints.The match had 34,520 attendees, one yellow card, two substitutions per team, and three refereeing officials of different nationalities.
  • Case 2a: DCI-Agent-CC on Agentic Search: A Guardian interview links David Platt’s goal to “technique” and connects Enzo Scifo with having performed it at Auxerre.These interview clues redirect the search from the match itself toward identifying the target club.
  • Case 2a: DCI-Agent-CC on Agentic Search: The final answer is AJ Auxerre, supported by identifying Enzo Scifo as the older of two Belgium players born in 1966 and by confirming his club history.Scifo was born 19 February 1966, Demol 11 March 1966, and the dates differ by exactly 20 days.

D.3 Case 2b: DCI-Agent-Lite on Agentic Search

DCI-Agent-Lite solves a multi-constraint BrowseComp-Plus query using only Bash and Read tools. It progressively narrows the corpus with targeted searches and verifies the final candidate by reading source documents.

  • Search process: Using only Bash and Read tools, DCI-Agent-Lite progressively narrows the search space through targeted rg commands despite its limited tool set.The trace includes searches for distinctive clues such as “scoreboard,” “principal,” and “tank.”
  • Result: The agent identifies the book as Made You Up by Francesca Zappia.The answer satisfies clues about publication in May 2015, the plot, and the author’s background.
  • Verification: The agent verifies the candidate by reading documents that confirm the lobster-tank plot detail, May 19, 2015 publication date, and the author’s early writing and drawing.It reads an interview with Francesca Zappia and the book’s Amazon document before producing the final answer.

D.4 Case 3a: DCI-Agent-CC on Knowledge-Intensive QA

This case shows DCI-Agent-CC performing single-hop factual retrieval on Natural Questions by searching and cross-referencing corpus documents. For a film-role question, it identifies Jehnna as the virgin in Conan the Destroyer and Olivia d’Abo as the actor.

  • Task scope: The case demonstrates single-hop factual retrieval on Natural Questions, a standard knowledge-intensive QA benchmark.The query concerns a specific film role and is answered through relevant Wikipedia articles.
  • Evidence synthesis: Cross-referencing the film article with the actor’s biography enables a confident, cited answer.The two documents independently connect Jehnna, Olivia d’Abo, and Conan the Destroyer.
  • Search process: The agent first locates the Conan the Destroyer article, then refines the search for Jehnna and Olivia d’Abo after finding incomplete context.This shows iterative corpus interaction using targeted lexical searches rather than relying on one truncated result.
  • Case result: DCI-Agent-CC identifies Jehnna, played by Olivia d’Abo, as the virgin in Conan the Destroyer.The final answer combines the Conan the Destroyer article with Olivia d’Abo’s biography as cited evidence.

D.5 Case 3b: DCI-Agent-Lite on Knowledge-Intensive QA · D.6 Case 4a: DCI-Agent-CC on Information Retrieval

The case studies show DCI agents solving knowledge-intensive QA and reasoning-intensive retrieval by directly searching local corpora, combining lexical discovery with targeted context reading. DCI-Lite extracts a precise count, while DCI-Agent-CC links distributed evidence to distinguish proximate mechanisms from ultimate explanations.

  • D.5 Case 3b: DCI-Agent-Lite on Knowledge-Intensive QA: DCI-Agent-Lite resolves a count query by inspecting corpus structure, disambiguating St. Louis Cardinals with football context, and extracting the relevant sentence.The corpus is a single large dump, so the agent progressively narrows from city and sports references to the likely city article.
  • D.5 Case 3b: DCI-Agent-Lite on Knowledge-Intensive QA: St. Louis has had 4 NFL teams: the St. Louis All-Stars, St. Louis Gunners, St. Louis Cardinals, and St. Louis Rams.The agent searched city and football references, then read neighboring corpus lines to recover the complete enumeration.
  • D.6 Case 4a: DCI-Agent-CC on Information Retrieval: The BRIGHT-Biology case requires joining phototaxis evidence with proximate-versus-ultimate causation evidence across two source-article groups.The identified gold documents include three Phototaxis files and two Proximate_and_ultimate_causation files.
  • D.6 Case 4a: DCI-Agent-CC on Information Retrieval: DCI-Agent-CC searches the biology directory with targeted terms including phototaxis, light stimulus, proximate, and ultimate causation, then reads the matched core documents.This workflow first locates candidate files and then separately inspects phototaxis and causation documents to answer the multi-part question.
  • D.6 Case 4a: DCI-Agent-CC on Information Retrieval: The retrieved phototaxis evidence defines directed movement toward increasing light and documents positive phototaxis in flying insects, including moths, grasshoppers, and flies.The passages characterize light, rather than a chemical, as the attractant and describe artificial lighting as a supernormal stimulus.
  • D.6 Case 4a: DCI-Agent-CC on Information Retrieval: The causation documents define proximate causes as immediate mechanisms and ultimate causes as evolutionary explanations, enabling the agent to interpret the heat-versus-LED argument correctly.The answer therefore treats the LED observation as evidence about the immediate cue, not the evolutionary origin of the behavior.
  • D.6 Case 4a: DCI-Agent-CC on Information Retrieval: The final answer identifies five relevant gold documents and concludes that low-heat LED attraction refutes heat as insects’ immediate cue, supporting light-driven phototaxis.It distinguishes proximate immediate mechanisms from separate ultimate evolutionary explanations rather than claiming that evolution is irrelevant.

D.7 Case 4b: DCI-Agent-Lite on Information Retrieval

DCI-Agent-Lite correctly identified 9745001.txt as the directly relevant SciFact paper supporting the radioiodine claim. It used targeted local-corpus searches and verification to distinguish the exact match from related thyroid and dosimetry papers.

  • D.7 Case 4b: DCI-Agent-Lite on Information Retrieval: The agent maximized recall with targeted searches for radioiodine, the exact disease phrase, and thyroid-volume reduction before ranking candidates.It searched only the local SciFact corpus and refined the query after finding several related thyroid papers.
  • D.7 Case 4b: DCI-Agent-Lite on Information Retrieval: It verified treatment, population, and outcome by reading the strongest candidate, while filtering out tangential multinodular-goitre documents.The final ranking contained 9745001.txt as the most relevant document.
  • D.7 Case 4b: DCI-Agent-Lite on Information Retrieval: 9745001.txt directly supports the claim: median thyroid volume fell from 73 ml to 29 ml at 24 months, a 60% reduction.The paper studied radioactive iodine in patients with non-toxic multinodular goitre and measured thyroid volume.

D.8 Case 5a: Failure Case of DCI-Agent-CC in Agentic Search

This case exposes a DCI-Agent-CC failure on a multi-hop BrowseComp-Plus football query: the agent identified Denis Suarez and FC Barcelona but ultimately answered Inter Milan instead of the correct FC Krasnodar.

  • Correct intermediate identification: The agent correctly linked the water-bottle incident to Denis Suarez and his loan from FC Barcelona to Arsenal.It found corpus evidence listing Denis Suarez at Barcelona and identified his January 2019 loan to Arsenal, but did not complete the final cross-document verification.
  • Failure outcome: The agent’s final answer was Inter Milan, whereas the correct team was FC Krasnodar.The supplied answer record identifies FC Krasnodar as correct, while the agent concluded that the bench appearance was against Inter Milan.
  • Search process: The failure arose despite extensive iterative searches across match reports, player profiles, club archives, and exact match-statistic clues.The agent used Grep and Bash searches for the incident, match statistics, loan records, and Barcelona’s season documents, but the multi-hop chain remained unresolved.
  • Reasoning limitation: The agent could not establish which European match included Denis Suarez on Barcelona’s bench and acknowledged insufficient corpus evidence.Its search established Barcelona’s Champions League opponents but did not pinpoint a specific bench appearance, leading to an unsupported final selection.

D.9 Case 5b: Failure Case of DCI-Agent-Lite in Agentic Search

This case shows DCI-Agent-Lite failing to solve a BrowseComp-Plus movie-identification task with multiple interlocking clues. Despite progressively refining its searches, the agent returned an incorrect answer after broad queries produced irrelevant results and failed to identify the correct corpus files.

  • Failure outcome: The agent struggled to formulate precise queries for clues spanning the 2000s release, contrasting backgrounds, actors born in the 1960s, and a November 2023 dispute.These interlocking clues required combining plot, actor, date, and director–actor-brother information.
Loading 2605.05242v1…