ReCite — Agentic Reasoning for Faithful Citation

Motivation

Similarity is not support

Topical relevance can point to an authentic paper that still fails to logically support the author’s claim. ReCite starts from the question: what does a citation actually need to prove?

Context determines evidence Background, comparison, and critique citations demand different verification criteria.
Intent should shape retrieval Keywords conditioned on citation intent outperform surface-level semantic matching.
Errors need a recovery loop A single faulty retrieval should trigger reflection, not propagate downstream.

Authentic databases only Strict F1 39.15%

Similarity-based retrieval versus the ReCite reasoning loop — **Figure 1.** Similarity-based retrieval vs. ReCite.

TL;DR

Citation requires logical support, not merely textual similarity. ReCite is a decoupled agentic framework that locates citation positions, plans intent-aware retrieval queries, and reflectively verifies whether each candidate supports the claim — re-retrieving when evidence is inconsistent. Built on Qwen3-4B with SFT and GRPO, ReCite achieves a Strict F1 of 39.15%, surpassing substantially larger generative models and general-purpose agents.

Method

A decoupled, closed-loop citation agent

ReCite turns citation into an observe–think–act workflow. Three specialized Qwen3-4B modules locate citation breakpoints, plan retrieval keywords, and verify candidates with reflective re-retrieval.

Figure 4. The decoupled ReCite architecture: CiteLocator, QueryPlanner, and Master Brain form an iterative loop of perception, planning, retrieval, verification, and reflection.

01

CiteLocator · SFT

Location perception

Predicts citation positions and labels each logical breakpoint Mandatory or Optional.

Overall F166.37%

02

QueryPlanner · SFT + GRPO

Intent-aware planning

Infers citation intent and generates targeted retrieval keywords per citation position.

Reasoning F157.81%

03

Master Brain · SFT

Verification & reflection

Orchestrates Semantic Scholar retrieval, verifies claim–evidence consistency, and re-queries on failure.

Strict F139.15%

Data

Supervision for reasoning, not just retrieval

Built from 10,893 LaTeX papers at top CS venues (2024–2025), the corpus provides three aligned subsets for location perception, intent reasoning, and reflective trajectories.

10,893 LaTeX papers ACL · EMNLP · CVPR · ICCV · NeurIPS

93,403 Citation labels CAP-8 intent taxonomy

31,118 Location samples 28,006 train / 3,112 test

2,000 Trajectories 500 gold + 1,500 variants

Conference and citation intent distributions — Figure 2. Venue and CAP-8 citation intent distributions of the ReCite corpus.

Reflective trajectory synthesis pipeline

Figure 3. Synthesized reflective trajectories teach the agent to diagnose retrieval drift and issue a refined secondary query.

Takeaway Vague queries and real distractors are injected on purpose — the agent must explain, correct, and re-query rather than memorize a single retrieval path.

Table 1. Dataset comparison

ReCite provides supervision beyond static context–reference pairs: it also records where citations appear, why they are needed, and how the agent should recover when retrieval fails.

Capability	RefSeer	S2ORC	ReCite
Local context mapping	✓	✓	✓
Full-text structure	✗	✓	✓
Location perception	✗	✗	✓
Intent reasoning	✗	✗	✓
Reflective trajectories	✗	✗	✓

What ReCite adds 3 signals beyond context→reference

01

Location

Where a citation is needed

02

Intent

Why a particular work is cited

03

Reflection

How to recover when retrieval drifts

Table 2. CAP-8 citation intent taxonomy

Each citation is annotated with one of eight intent categories, which drives QueryPlanner’s retrieval-keyword generation and the verification criteria of the Master Brain.

Abbr.	Category	Description
BG	Background	Provides background knowledge or information for the research domain.
MG	Motivation & Gap	Highlights research motivations or identifies gaps in existing literature.
CMP	Comparison	Compares the current work or other baselines with specific prior methods.
BS	Basis & Support	Provides theoretical foundations or concepts supporting the current study.
MTD	Method & Data	Refers to specific methods, algorithms, tools, or datasets utilized in the research.
VLD	Validation	Provides empirical evidence, results, or evaluation metrics to validate claims.
EV	Evolution	Describes the historical evolution or development of a methodology.
CRT	Critique	Expresses critiques, conflicts, or opposing views regarding prior work.

Results

Main results

Evaluated on 200 held-out paragraphs and 1,023 unseen target papers from 2023–2025 top-venue submissions.

39.15 Strict F1 (%) best among all methods

55.14 Lenient F1 (%) logical alternatives found

89.71 Position-only F1 (%) exact placement accuracy

4B Model size rivals 1.6T-scale baselines

Figure 5. Recall across the CAP-8 citation intent categories.

Takeaway ReCite keeps higher recall than most baselines across intent categories, showing that explicit intent supervision aligns retrieval with the author’s motivation rather than surface overlap.

Table 3. End-to-end citation prediction

End-to-end workflows are scored at three strictness levels: exact target match (Overall-Strictly), any logically supportive alternative (Lenient), and citation position only. ReCite-SFT (CAP-8) achieves the best Strict F1 even against 1.6T-scale generative baselines.

Method	Size	Overall-Strictly			Lenient Evaluation			Position-Only
Method	Size	P↑	R↑	F1↑	P↑	R↑	F1↑	P↑	R↑	F1↑
Zero-shot prompt pipelines
GLM-4-Plus	–	9.83	10.27	10.05	17.22	17.99	17.60	47.08	49.19	48.11
Mimo-V2.5-Pro	–	26.40	27.72	27.04	29.19	30.63	29.89	75.07	78.78	76.88
GPT-4o-mini	–	7.20	8.65	7.86	12.73	15.29	13.89	44.89	53.90	48.98
GPT-5.1-Chat	–	8.99	12.12	10.32	14.48	19.54	16.63	35.32	47.64	40.57
Kimi-k2-preview	–	19.06	21.24	20.09	24.32	27.10	25.64	55.86	62.24	58.88
Qwen3.6-Plus	–	28.09	35.44	31.34	31.09	39.23	34.69	68.60	86.50	76.52
Qwen3.6-27B	27B	22.89	29.62	25.83	25.58	33.08	28.85	62.34	80.48	70.26
DeepSeek-V4-Flash	284B	31.23	35.21	33.10	34.52	38.89	36.57	72.95	82.11	77.26
DeepSeek-V4-Pro	1.6T	31.29	32.28	31.77	35.25	36.34	35.79	75.97	78.32	77.13
General agents (end-to-end via Mimo-V2.5-Pro)
Claude Code	–	17.73	27.26	21.49	21.35	32.82	25.87	25.51	39.23	30.92
Hermes Agent	–	22.39	21.24	21.80	26.38	25.02	25.68	28.01	26.56	27.27
OpenClaw	–	18.09	7.03	10.12	25.25	9.81	14.13	29.42	11.43	16.46
ReCite (Ours)
ReCite-Base	4B	36.92	35.75	36.33	43.46	42.08	42.76	89.80	86.11	87.53
ReCite-SFT	4B	37.48	37.45	37.47	47.14	47.07	47.10	89.80	89.66	89.73
ReCite-SFT (CAP-8)	4B	39.22	39.07	39.15	55.27	55.02	55.14	89.92	89.51	89.71

P / R / F1 scores (%). Bold blue marks the best value in each column; shaded rows are ReCite models. CAP-8 denotes fine-tuning with citation taxonomy information.

Figure 6. Cross-venue robustness across EMNLP, CVPR, ICCV, and NeurIPS test paragraphs.

Figure 7. Performance vs. the number of retrieved candidates. k = 8 balances evidence sufficiency and latency.

Takeaway Performance is stable across venues, and k = 8 provides the best evidence/latency trade-off — more candidates add context noise rather than accuracy.

Evaluation

Module-level results and analysis

Each module is benchmarked separately against strong LLM and heuristic baselines, followed by a full ablation of the decoupled design.

Table 4. Citation location prediction (N = 3,112)

CiteLocator must decide not only where a citation belongs, but whether it is Mandatory or Optional. It outperforms every LLM baseline in Overall F1 and narrows the gap to human-level placement.

Model	Size	Strategy	Mandatory			Optional			Overall
Model	Size	Strategy	P↑	R↑	F1↑	P↑	R↑	F1↑	P↑	R↑	F1↑
Baseline models
GPT-5.1-Chat	–	Direct	22.01	59.72	32.17	9.21	58.50	15.92	27.73	59.39	37.81
GPT-4o-mini	–	Direct	24.62	53.61	33.75	11.34	57.17	18.92	31.25	54.57	39.74
Kimi-k2-preview	–	Direct	31.27	61.05	41.36	14.36	61.28	23.27	38.38	61.11	47.15
MiniMax-M2.5	–	Direct	33.52	56.37	42.04	14.71	52.52	22.98	40.35	55.34	46.67
GLM-4-flash	–	Direct	25.81	16.73	20.30	12.07	17.96	14.43	32.67	17.06	22.41
Qwen3-Max	–	Direct	32.87	28.90	30.75	15.41	29.28	20.19	40.18	29.00	33.69
Qwen3-Max	–	CoT	34.43	28.42	31.14	16.00	28.09	20.39	41.71	28.33	33.74
DeepSeek-V3	671B	Direct	30.91	62.66	41.40	13.91	61.65	22.70	37.85	62.39	47.12
DeepSeek-V3	671B	CoT	29.23	49.09	36.64	12.41	45.88	19.54	35.68	48.23	41.02
Qwen3.5-27B	27B	Direct	36.13	55.86	43.88	15.83	50.56	24.11	42.98	54.44	48.04
Qwen3-4B	4B	Direct	13.98	9.70	11.45	5.55	9.56	7.03	18.12	9.66	12.60
Qwen3-4B	4B	CoT	8.98	3.99	5.52	2.59	2.92	2.74	11.12	3.70	5.55
ReCite
CiteLocator	4B	Direct	67.93	60.61	64.06	42.90	58.57	49.52	74.16	60.06	66.37

P / R / F1 scores (%). CiteLocator reports Precision, Recall, and F1 for Mandatory, Optional, and Overall categories.

Table 5. Intent-aware query planning (N = 300)

QueryPlanner is evaluated on the quality of its reasoning trace and the keywords it extracts. GRPO lifts reasoning F1 close to DeepSeek-V3’s level while using only 4B parameters.

Model	Size	Reasoning		Keyword extraction
Model	Size	R-F1↑	R-Jdg↑	K-F1↑	K-TR↑	K-PR↑	K-Jac↑	K-Jdg↑
Baseline models
GPT-5.1-Chat	–	52.25	7.48	39.05	44.03	12.16	7.87	6.26
MiniMax-M2.5	–	53.88	7.86	41.19	47.19	13.73	8.75	6.53
Qwen3-Max	–	53.91	7.33	37.01	36.56	15.38	10.83	5.95
GLM-4-flash	–	21.73	3.04	16.09	19.95	6.65	3.98	2.63
Kimi-k2-preview	–	54.12	7.75	45.80	55.30	25.45	16.42	7.12
DeepSeek-V3	671B	58.63	7.88	46.25	49.18	22.59	15.13	6.65
Qwen3.5-27B	27B	50.55	7.56	40.99	43.62	16.73	11.14	6.33
ReCite
QueryPlanner (SFT)	4B	56.12	7.21	37.74	43.56	16.87	10.38	5.93
QueryPlanner (SFT + GRPO)	4B	57.81	7.64	39.06	42.47	17.85	11.00	6.44

Reasoning R-F1: token-level F1 of the generated reasoning; R-Jdg / K-Jdg are LLM-as-a-judge scores. Keyword metrics cover token F1 (K-F1), token recall (K-TR), phrase recall (K-PR), and Jaccard similarity (K-Jac).

Table 6. BUSCALDI out-of-domain benchmark

On an unseen citation-location benchmark, CiteLocator generalizes beyond CS-specific rhetorical signals and beats heuristic baselines on fine-grained placement.

Type	Method	P↑	R↑	F1↑
Coarse-grained
	Scientist	76.2	88.4	81.8
	GM-s2orc	78.2	78.2	78.2
	GM-s2orc-H	80.2	82.6	81.4
	CiteLocator	66.7	87.0	75.5
Fine-grained
	Scientist	57.5	66.6	61.7
	GM-s2orc	44.9	44.9	44.9
	GM-s2orc-H	49.2	50.7	50.0
	CiteLocator	46.7	60.9	52.8

Values are percentages. “-H” indicates extra heuristic rules.

Table 7. Ablation of ReCite components

Each module is individually replaced by the Qwen3-4B-Instruct baseline. Location perception is foundational; intent-aware planning is indispensable; reflective verification mainly improves Lenient F1.

Method	Overall-Strictly			Lenient Evaluation
Method	P	R	F1	P	R	F1
ReCite	37.48	37.45	37.47	47.14	47.07	47.10
w/o CiteLocator	2.98	2.55	2.75	7.05	6.02	6.49
w/o QueryPlanner	15.79	15.75	15.77	30.88	30.81	30.85
w/o Master Brain	36.92	35.75	36.33	43.46	42.08	42.76

Replaced modules use the Qwen3-4B-Instruct baseline. Removing CiteLocator or QueryPlanner causes severe drops; removing Master Brain’s reflective verification mainly hurts Lenient F1.

Citation

BibTeX

@inproceedings{huang2026recite,
  title     = {ReCite: Agentic Reasoning for Faithful Citation},
  author    = {Huang, Yuyang and Li, Bobo and Song, Jiajia and Ding, Yuzhe and Teng, Chong and Li, Fei and Ji, Donghong},
  booktitle = {Findings of the Association for Computational Linguistics: EMNLP 2026},
  year      = {2026}
}