TrendR: A Recoverable Multi-Agent Literature Review Framework for Research Scenarios

TrendR transforms the traditional one-off literature review process into a recoverable, traceable, and verifiable control-plane systemIt manages the research process via an explicit state machine, addressing core pain points such as multi-round retrieval, insufficient analysis depth, and resuming from breakpoints.

In academic research, literature review is an essential foundational taskHowever, traditional literature review tools often use a "one-off generation" thin-layer prompt workflow: searching, summarizing, and drafting are done in one goThis model has several obvious flaws:

First, Irrecoverable process: Once the operation is interrupted, it often needs to be restarted from scratch, and previous work results are difficult to inherit effectivelySecond, Weak quality control: Generation and verification are usually self-assessed by the same agent, lacking an independent verification mechanismFinally, Limited scalability: Loose intermediate text dependencies make it difficult to migrate and reuse the process across multiple platforms.

The TrendR project is designed to address these pain pointsIt aims to transform literature review from a one-off generation-dominated model into a controlled research pipeline with governance state, product contracts, independent verification, and recoverable execution.

TrendR's core identity is a recoverable literature review harness (framework) rather than a simple writing toolIts design philosophy can be summarized into four key transformations:

Traditional workflows rely on loose intermediate texts, while TrendR uses an explicit state machine to manage the research processThe core path is clearly defined as: INIT → DISCOVERY → ANALYSIS → GAP_CHECK → WRITING → VERIFY → DONEEach step is determinable, traceable, and recoverableThe process is no longer "finished once written" but progresses by state, with backtracking and controlled retries.

Phase transitions are based on structured files rather than context guessingTypical products include:

• candidates.csv: Candidate paper pool
• matrix.csv: Structured analysis matrix
• gap_report.md: Coverage gaps and backtracking basis
• review.md: Review main text
• verify.json: Independent verification results
• run_state.json: Machine-readable running state

These file contracts give the process engineering boundaries that are recoverable, debuggable, and reproducible.

TrendR separates generation and verificationWhether the task is completed is determined by the verifier, not self-declared by the writing agentThe verifier focuses on three core checks: citation consistency, claim support, and taxonomy coherenceTherefore, "completion" is an externally determined result, not a subjective conclusion attached to the generation process.

The system retains machine-readable state and heartbeat, allowing recovery and observation based on run_state and heartbeat informationRuntime-related logic is isolated in the adapter layer, and the core state machine is not coupled to a single platform, enabling the same control logic to be reused across runtimes.

According to the project documentation, the measured limits of TrendR's current pipeline are as follows:

Metric	Actual Reachable Limit	Theoretical Absolute Limit	Source of Limitation
Discovery Rounds	6 rounds	12 rounds (×resume)	State machine hardcoding
Single Run Duration	~5 hours	60 hours (12 resumes)	Cumulative state timeout
Page Visits	600–1000 times	~2000 times	Rounds × Sources
Candidate Paper Pool	650 papers	~1000 papers	Convergence after deduplication
Actually Analyzed Papers	80 papers (Depth C)	80 papers (target_papers bottleneck)	CLI configuration value

The current pipeline has three key breakpoints:

1** Analysis capacity is far less than collection capacity**: The collection limit is 650 papers, while the analysis limit is only 80 papers, an 8-fold gap; about 500 papers with scores 3–4 are directly discarded, lacking secondary indexing

2** Depth C round conflict**: cli.py allows --max-rounds 10, but state_machine.py has a hardcoded upper limit of 6, making the 10-round design unreachable

3** Zotero integration is still a stub**: references.bib is generated but needs manual import; cross-project indexing for paper-pool.csv exists but has no visualization; the paper "library" is available, but the relationship network between nodes (co-citation, topic clustering) has not been established.