PaSaMaster: Agentic Lit Retrieval

• PaSaMaster is a self-evolving, evidence-grounded literature retrieval system that reframes search as iterative intent and relevance ranking.
• It separates planning from retrieval using a Navigator and Librarian Swarm, ensuring verified metadata and zero source hallucination.
• The system outperforms traditional engines with significant F1 improvements and cost-efficient, auditable search processes.

to=arxiv_search.search ＿老司机_code {"query":"PaSaMaster self-evolving agentic literature retrieval (Du et al., 14 May 2026)", "max_results": 5} maybe need exact function name. to=arxiv_search.search 神彩争霸 一级a做爰片 {"query":"(Du et al., 14 May 2026) PaSaMaster", "max_results": 10} to=arxiv_search 天天种彩票_code {"query":"(Du et al., 14 May 2026) PaSaMaster", "max_results": 10} to=arxiv_search.search 手机上天天中彩票json {"query":"agentic literature retrieval benchmark hallucination scientific search", "max_results": 5} PaSaMaster is a self-evolving, evidence-grounded agentic literature retrieval system that produces relevance-scored paper rankings with evidence-grounded recommendations through iterative intent analysis, retrieval, and ranking. It was introduced to address a twofold challenge in scientific search: ensuring source authenticity while maintaining a deep comprehension of academic search intents. The system reframes literature discovery as intent–paper relevance ranking rather than generation, separates planning from retrieval, and uses ranked evidence to refine search intent and strategy over multiple rounds rather than treating retrieval as one-shot query–document matching (Du et al., 14 May 2026).

1. Conceptual framing and motivation

PaSaMaster is positioned against three established retrieval paradigms. Traditional keyword-centric engines preserve authenticity by returning only indexed records, but they compress rich academic intents into short queries and thereby lose constraints such as methods, datasets, temporal ranges, or exclusion rules. Dense semantic retrieval improves lexical matching, yet remains passive one-shot query–document matching and still compresses complex intents. Generative LLM agents, by contrast, can understand nuanced natural-language intents, but remain vulnerable to hallucinated sources, inconsistent citation fields, and high compute cost (Du et al., 14 May 2026).

The system’s central reformulation is to treat literature retrieval as intent–paper relevance ranking rather than generation. In this formulation, search is not a single retrieval act but an iterative process in which ranked evidence reveals gaps, refines intents, and guides follow-up searches. This design is described as transforming retrieval from a one shot query–document matching problem into a search process that evolves over time. The paper identifies three key designs: iterative self-evolving search, hallucination prevention through ranking real papers rather than generating citations, and cost efficiency through planning–retrieval separation (Du et al., 14 May 2026).

A common misconception is to treat PaSaMaster as merely another semantic search engine or as a citation-producing agent. The paper argues against both interpretations. It is not a purely lexical or dense retrieval backend, because its output depends on explicit intent decomposition, checklist construction, evidence localization, and reflection. It is also not a generative citation system, because the Navigator never outputs papers directly; all recommended papers must come from the repository and be accompanied by verified metadata and evidence (Du et al., 14 May 2026).

2. Formal problem statement

PaSaMaster defines a customized scientific repository $D$ with three tiers,

$$D = \{D_{\text{meta}}, D_{\text{abs}}, D_{\text{chunk}}\},$$

where $D_{\text{meta}}$ stores structured metadata, $D_{\text{abs}}$ stores abstract-level embeddings, and $D_{\text{chunk}}$ stores passage-level chunks from full texts. Agent tools are partitioned as $T = T_{\text{ret}} \cup T_{\text{read}}$, distinguishing retrieval from reading operations. Given a natural-language query $q$, the system induces an explicit intent representation $I$ that contains a retrieval plan $S$ and a verification checklist $C = [c_1,\dots,c_m]$, where each checkpoint $$D = \{D_{\text{meta}}, D_{\text{abs}}, D_{\text{chunk}}\},$$0 is an objective criterion a relevant paper must satisfy (Du et al., 14 May 2026).

For a paper $$D = \{D_{\text{meta}}, D_{\text{abs}}, D_{\text{chunk}}\},$$1 and checkpoint $$D = \{D_{\text{meta}}, D_{\text{abs}}, D_{\text{chunk}}\},$$2, PaSaMaster localizes evidence chunks $$D = \{D_{\text{meta}}, D_{\text{abs}}, D_{\text{chunk}}\},$$3 from $$D = \{D_{\text{meta}}, D_{\text{abs}}, D_{\text{chunk}}\},$$4 and assigns a checkpoint score $$D = \{D_{\text{meta}}, D_{\text{abs}}, D_{\text{chunk}}\},$$5 plus an evidence-grounded rationale. The per-paper criterion score is

$$D = \{D_{\text{meta}}, D_{\text{abs}}, D_{\text{chunk}}\},$$6

The Scorer also outputs a calibrated probability $$D = \{D_{\text{meta}}, D_{\text{abs}}, D_{\text{chunk}}\},$$7 for overall relevance. The final per-paper relevance score is

$$D = \{D_{\text{meta}}, D_{\text{abs}}, D_{\text{chunk}}\},$$8

A listwise reranker then performs global cross-paper comparison to produce a ranked list. The generic representation is written as

$$D = \{D_{\text{meta}}, D_{\text{abs}}, D_{\text{chunk}}\},$$9

where $D_{\text{meta}}$0 encodes the intent checkpoints, $D_{\text{meta}}$1 encodes paper metadata, abstracts, and evidence, and $D_{\text{meta}}$2 is a learned scorer. In PaSaMaster, $D_{\text{meta}}$3 is a lightweight model distilled from a stronger teacher, with evidence-grounded supervision (Du et al., 14 May 2026).

The training objective for the distilled Scorer combines checklist-level supervision with holistic relevance calibration:

$D_{\text{meta}}$4

where $D_{\text{meta}}$5 are checklist labels, $D_{\text{meta}}$6 is overall relevance, and $D_{\text{meta}}$7 balances calibration. This formalization makes the checklist the primary machine-checkable representation of search intent, rather than a latent by-product of a generative prompt (Du et al., 14 May 2026).

3. Architecture and self-evolving workflow

The architecture is explicitly divided between a single high-capability “Navigator” LLM and a parallel “Librarian Swarm.” The Navigator performs intent analysis, disambiguation, decomposition, and self-reflection. The Librarian Swarm executes large-scale retrieval, evidence localization, checklist scoring, and reranking over the repository. This planning–retrieval separation is the mechanism by which PaSaMaster reserves expensive frontier-model computation for intent understanding while delegating large-scale search and relevance scoring to customized corpora and lightweight models (Du et al., 14 May 2026).

The core loop begins with planning. The Navigator parses the query into $D_{\text{meta}}$8 and $D_{\text{meta}}$9, where $D_{\text{abs}}$0 describes sources and operators to use and $D_{\text{abs}}$1 encodes verifiable constraints such as task scope, method class, dataset, population, timeframe, and inclusion or exclusion rules. It may also generate an adaptive search plan including web-grounded disambiguation when needed. Retrieval then proceeds through multiple channels, after which candidates are scored against the checklist and reranked globally (Du et al., 14 May 2026).

The iterative character of the system is formalized through the update of the intent state. Let $D_{\text{abs}}$2, and let $D_{\text{abs}}$3 be ranked evidence extracted in round $D_{\text{abs}}$4. The Navigator updates

$D_{\text{abs}}$5

where $D_{\text{abs}}$6 summarizes uncovered constraints, distribution of satisfied and unsatisfied checkpoints, and missing topical or methodological strata observable from $D_{\text{abs}}$7. Reflection inspects the ranked set, identifies gaps such as missing subareas, underspecified constraints, or under-covered time windows, and refines both strategy and checklist (Du et al., 14 May 2026).

The paper emphasizes three concrete mechanisms by which ranked evidence guides later searches. Coverage charts compute per-checkpoint satisfaction histograms and topic or method coverage over top-$D_{\text{abs}}$8 results to detect gaps. Constraint sharpening rewrites ambiguous checklist items explicitly and can promote optional constraints if gaps persist. Strategy switching changes retrieval channels, for example by emphasizing citation expansion around a newly identified seminal work or pivoting to a subfield-specific repository. This suggests that PaSaMaster operationalizes search as an auditable control loop rather than as a static retrieval query (Du et al., 14 May 2026).

4. Retrieval, evidence localization, and anti-hallucination design

PaSaMaster uses three complementary retrieval channels: Semantic Direct Retrieval, Citation Network Expansion, and Web-to-Repository Verification. Candidate generation forms the union

$D_{\text{abs}}$9

Semantic Direct Retrieval derives query-side representations from $D_{\text{chunk}}$0 and $D_{\text{chunk}}$1, uses $D_{\text{chunk}}$2 for abstract-level filtering, and uses $D_{\text{chunk}}$3 for chunk-level evidence via cosine similarity,

$D_{\text{chunk}}$4

Citation Network Expansion expands seed candidates via cited and citing graphs to surface structurally proximate literature. Web-to-Repository Verification uses web search only to propose candidates; each proposed item must be mapped to $D_{\text{chunk}}$5 before eligibility (Du et al., 14 May 2026).

Evidence localization is checkpoint-specific. For each candidate paper $D_{\text{chunk}}$6 and checkpoint $D_{\text{chunk}}$7, the system retrieves supporting passages through a shared encoder:

$D_{\text{chunk}}$8

The Scorer then assigns $D_{\text{chunk}}$9 and an evidence-grounded rationale, records $T = T_{\text{ret}} \cup T_{\text{read}}$0, and computes $T = T_{\text{ret}} \cup T_{\text{read}}$1 and $T = T_{\text{ret}} \cup T_{\text{read}}$2. After per-paper scoring, a listwise model compares candidates jointly to refine the ranking. The paper notes that while it does not commit to a particular loss, a standard listwise softmax objective on graded gains can be used for the reranker (Du et al., 14 May 2026).

The anti-hallucination guarantees are design-level rather than post hoc. The Navigator never outputs papers directly. All recommendations must come from $T = T_{\text{ret}} \cup T_{\text{read}}$3 and be accompanied by evidence chunks and metadata verified against $T = T_{\text{ret}} \cup T_{\text{read}}$4. Any web-found candidate must be matched to a repository record before eligibility. Each paper is returned with Score + Evidence + Reason at each checkpoint, and missing evidence disqualifies satisfaction. Citation fields such as title, author list, venue, year, and link are pulled from $T = T_{\text{ret}} \cup T_{\text{read}}$5, not generated; deduplication merges variants across identifiers. The paper summarizes this as “zero source hallucination” (Du et al., 14 May 2026).

5. Benchmark, metrics, and empirical results

Evaluation is conducted on PaSaMaster-Bench, a benchmark of 244 expert-curated tasks across 38 disciplines spanning AI & Computing, Engineering & Technology, Medicine & Life Sciences, Basic Sciences, and Others. Queries are realistic natural-language intents with multiple constraints involving methods, datasets, populations, timeframes, publication filters, and inclusion or exclusion criteria. For each query, experts provide a checklist $T = T_{\text{ret}} \cup T_{\text{read}}$6; candidate sets are collected from multiple channels, and only papers satisfying all required checkpoints enter the target set $T = T_{\text{ret}} \cup T_{\text{read}}$7. Systems return $T = T_{\text{ret}} \cup T_{\text{read}}$8 up to $T = T_{\text{ret}} \cup T_{\text{read}}$9, and evaluation is reported at $q$0 using Precision, Recall, F1, and NDCG (Du et al., 14 May 2026).

The benchmark results reported in the paper are as follows:

System	F1@20	Hallucination / Cost
PaSaMaster	21.69%	0%, $0.05/query
Google Scholar	1.39%	lexical baseline
OpenScholar	7.92%	semantic baseline
Bohrium Science Navigator	12.26%	semantic baseline
Google Scholar Labs	18.87%	fixed-pipeline agent
GLM-5	18.18%	29.07%, $0.56
GPT-5.2	16.69%	11.80%, $6.06
Gemini-3.1-pro	12.48%	32.41%, $0.38
Kimi-K2.5	17.36%	35.67%, $0.16
DeepSeek-v3.2	15.56%	20.57%, $0.28
MiniMax-M2.7	15.11%	37.79%, $0.18

PaSaMaster attains $q$1, $q$2, $q$3, and $q$4, with a hallucination rate of $q$5 and a cost of $q$6 per query. The paper states that this improves F1 by $q$7 over Google Scholar’s $q$8, exposes hallucination rates up to $q$9 in generative LLM agents, and outperforms GPT-5.2 by $I$0 in $I$1 while using approximately $I$2 of GPT-5.2’s computational cost (Du et al., 14 May 2026).

Per-discipline robustness is also emphasized: PaSaMaster achieves the highest or comparable F1 across all 38 disciplines. Qualitative ablations further report that disabling reflection degrades F1 relative to iterative refinement; omitting citation expansion reduces recall for niche subcommunities; omitting web-to-repository verification reduces candidate diversity for newly emerging topics; removing listwise comparison can hurt NDCG; and additional rounds generally improve coverage until saturation, after which diminishing returns appear. This suggests that the self-evolving loop and the heterogeneous retrieval backend are both integral rather than optional (Du et al., 14 May 2026).

6. Reliability, practical usage, and limitations

The system’s reliability claims are grounded in architecture and evaluation. No paper enters the ranking without being present in $I$3, and all metadata fields are drawn from $I$4. Each checkpoint requires supporting evidence snippets $I$5, and failures to locate evidence reduce $I$6 and therefore $I$7. Deduplication combines identifier-based and metadata-fuzzy matching, while citation consistency checks perform cross-field validation such as DOI-title consistency prior to final output. On PaSaMaster-Bench, these mechanisms yield zero source hallucination across 244 tasks (Du et al., 14 May 2026).

The implementation is open source at https://github.com/sjtu-sai-agents/PaSaMaster. The described workflow consists of setup, configuration, execution, and extensibility. Setup includes installing dependencies, preparing access to a frontier LLM for the Navigator, selecting a lightweight Scorer checkpoint, and building or pointing to a corpus $I$8 by ingesting metadata, generating abstract embeddings in $I$9, and chunking full texts into $S$0. Configuration enables retrieval tools, sets budgets for rounds and top-$S$1 retrieval per channel, and chooses planner prompts for PLAN and REFLECT to emit both $S$2 and $S$3 in structured form. Running the system returns a ranked list with, for each paper, score $S$4, metadata, and checkpoint-level Score + Evidence + Reason (Du et al., 14 May 2026).

The paper identifies several limitations and future directions. Specialized domains such as rare subfields and non-English corpora may benefit from domain-tuned encoders and Scorers. Multilingual support requires extending $S$5 and $S$6 encoders and Navigator prompts. Better intent modeling could use richer semi-formal query schemas such as PICO for clinical queries. Human-in-the-loop feedback could help accept or refine $S$7 and steer $S$8 for hard queries. Freshness and coverage of the repository remain operational concerns in rapidly evolving areas, and proprietary corpora require on-prem deployment, secure storage of $S$9, and strict tool access controls (Du et al., 14 May 2026).

The resulting system is particularly suited to settings where source authenticity and full-intent satisfaction matter, including systematic reviews, grant scouting, and method audits. The paper’s broader implication is that literature retrieval can be made more faithful to scientific intent by replacing citation generation with evidence-grounded ranking, and by turning search into an iterative, auditable process whose internal state is explicit in the pair $C = [c_1,\dots,c_m]$0 (Du et al., 14 May 2026).