Faithful citation attribution begins with identifying the intended source for a scientific claim. We study this source-identification capability through citation recovery: recovering the paper cited by the original author from a citation-bearing passage. Our evaluation adopts the published author's citation as an observable human attribution signal and uses target recovery as a proxy for progress toward faithful attribution. We introduce ATTRICITE, an open 4B-parameter model trained for tool-using citation recovery within the CiteGuard retrieval environment, together with CITEALIGN, a 7,607-instance computer-science dataset drawn from recent scientific literature. For controlled evaluation, we construct a 709-instance benchmark subset of CITEALIGN, comprising 410 development instances from 2024 publications and 299 temporally held-out test instances from 2025 publications. Across three runs at an inference temperature of 0.7, GRPO fine-tuning improves Qwen3-4B from 49.4%±1.5% to 59.8%±0.2% target-match accuracy, a gain of 10.4 percentage points. Despite using only 4B parameters, ATTRICITE outperforms gpt-oss-20b and comes within 3.9 points of GPT-5.4-mini, while Gemma 4 31B IT achieves the strongest overall performance at 72.0%±1.0%. We release the model and collection pipeline https://github.com/KathCYM/AttriCite to support reproducible research on citation recovery toward faithful attribution in a continually evolving scientific literature.
Figures & tables
Figure 1: Open-ended citation recovery with AttriCite . Given a passage with a missing citation, the model searches an external scholarly index and identifies the source cited by the original author among papers with similar terminology.
Table 2
Figure 2: Validation accuracy across training checkpoints. Step 475 is selected for test evaluation.
Figure 3: Target-match accuracy on (a) CiteAlign and (b) the biomedical cross-domain set. In both panels, the deterministic construction baseline submits the top-ranked Semantic Scholar result for a fixed query comprising the first eight non-stopword passage tokens and has no error bar. Biomedical runs use the Medicine field filter. All results use identifier-first entity matching with the frozen title fallback. Model error bars show two-sided 95% t -intervals over three runs; Claude Haiku 4.5 is evaluated once.
System
Accuracy
SD
95% CI
First-8 query top-1
35.45
–
–
Qwen3-4B (base)
49.39
1.54
45.55–53.22
gpt-oss-20b
49.72
1.84
45.15–54.30
Gemma 4 31B IT
72.02
0.97
69.62–74.42
GPT-5.4-mini
63.66
1.84
59.08–68.23
Claude Haiku 4.5 (1 run)
32.78
–
–
Table 3: Target-match accuracy on CiteAlign at temperature 0.7. The deterministic construction baseline uses the top-ranked result for a fixed first-eight-content-token query. For models, we report three-run means, standard deviations, and two-sided 95% t -intervals, except for Claude Haiku 4.5, which is evaluated once because of API cost.
Model
Accuracy
SD
95% CI
First-8 query top-1
25.87
–
–
Qwen3-4B (base)
42.89
1.07
40.24–45.54
AttriCite (ours)
54.08
3.15
46.25–61.91
Table 4: Target-match performance on the 143-instance biomedical cross-domain set using Semantic Scholar’s Medicine field filter and the same identifier-first entity matcher with frozen title fallback used for the main set. The deterministic First-8 baseline is a single frozen run; model results are summarized across three runs.
Model
Retrieval recall
Selection ∣ retrieved
Accuracy
First-8 query top-1
35.45
–
35.45
Qwen3-4B (base)
58.64
84.22
49.39
AttriCite (ours)
68.12
87.73
59.75
GPT-5.4-mini
70.90
89.78
63.66
Table 5: Retrieval and selection performance pooled across three runs at temperature 0.7. The deterministic baseline uses its single top-ranked result as the selector, so retrieval at rank one and correct selection coincide.
Appendix figures & tables4 assets
Supplementary material from the paper’s appendix.
Appendix
Configuration
Value
Base model
Qwen3-4B
Optimization
Full-parameter GRPO
Training epochs
3
Rollouts per prompt
8
Prompt batch size
2
Learning rate
1×10−6
Appendix
Table 6: AttriCite training configuration. We optimize all Qwen3-4B parameters with GRPO under the same five-action budget used at evaluation.
Training set
Reward
Accuracy
Retrieval recall
Selection ∣ retrieved
None
Base model
49.39±1.54
58.64
84.22
Subset
Retrieval-shaped
51.95±0.19
62.54
83.07
Subset
Outcome
59.75±0.19
68.12
87.73
1,000
Retrieval-shaped
58.31±2.18
67.22
86.73
1,000
Outcome
58.97±0.84
68.67
85.88
Appendix
Table 7: Exploratory outcome- versus target-retrieval-shaped reward comparison on the 299-instance main test set. Accuracy is the mean across three temperature-0.7 runs with sample standard deviation. Retrieval recall and selection conditional on retrieval are pooled across the same runs. All quantities use the identifier-first evaluator with the frozen title fallback.
Reported name
Exact identifier
License or terms
Qwen3-4B
Qwen/Qwen3-4B
Apache 2.0
AttriCite (ours)
step-475 derivative of Qwen3-4B
Apache 2.0
gpt-oss-20b
openai/gpt-oss-20b
Apache 2.0
Gemma 4 31B IT
google/gemma-4-31B-it
Apache 2.0
GPT-5.4-mini
gpt-5.4-mini-2026-03-17
OpenAI API terms
Claude Haiku 4.5
claude-haiku-4-5-20251001
Anthropic API terms
Appendix
Table 8: Model and service identifiers used in the experiments. Hosted models are governed by the provider terms listed here rather than an open-weight license.
System
Identifier
Exact title
Positive override
Negative override
Other incorrect
Qwen3-4B (base), 3 runs
430
5
8
21
433
AttriCite , 3 runs
515
12
9
16
345
gpt-oss-20b, 3 runs
432
4
10
23
428
Gemma 4 31B IT, 3 runs
618
11
17
7
244
GPT-5.4-mini, 3 runs
547
13
11
26
300
Claude Haiku 4.5, 1 run
94
1
3
3
198
Appendix
Table 9: Entity-matching pathways aggregated over the primary main-table test runs. “Exact title” applies only after identifier mismatch; positive and negative overrides are manually adjudicated ambiguous pairs. Every fuzzy-only pair credited in these primary runs was reviewed.