KnowMe Paper Radar: Lightweight Discovery of Promising Papers in the Age of Autoresearch
Organizations: Ant Group
Abstract
The growing volume of topically similar papers makes it difficult to identify new work that can offer useful research insights or methodological guidance. We present KnowMe Paper Radar, a lightweight discovery tool that combines multiple sources of evidence into personalized reading priorities. Users specify topics, priorities, exclusions and dates; the system brings together topical relevance, recency, author research histories, collaboration paths and direction-specific institutional context. By exposing each score component and its supporting sources, KnowMe helps readers select promising papers for their current questions and discover relevant work beyond familiar authors. Automatic acquisition, version merging and graph-cache reuse keep this workflow convenient without large language model (LLM) inference during ordinary use. Experiments on 174,640 acquired records show more than threefold faster repeated context preparation, while cached abstracts improve normalized discounted cumulative gain at rank 10 (NDCG@10) from 0.412 to 0.598 on expert-judged retrieval tasks. Complementary analyses examine how the same accumulated records support semantic retrieval, research-history modeling and literature-grounded planning. A browser interface and a structured command-line interface (CLI) make these capabilities available to both researchers and research agents, positioning KnowMe as a practical upstream tool for literature-informed recursive self-improvement (RSI) workflows.
Figures & tables
| Level | Measured operation | Observed result |
| Paper / version | Collapse repeated eligible ID/version occurrences | 44,512 repeats merged |
| Source response | Re-filter 142 cached pages under an expanded scope | 1,533 IDs recovered |
| Source traversal | Reuse raw and structured pages from 508 processed pages | 282 read locally |
| Metadata repair | Consolidate daily repair windows into a bounded plan | 945 108 windows |
| Canonical library | Merge first submissions and earlier updated papers | 174,640 in-scope IDs |
| Direction | Actual records | Fresh source (s) | Retained source (s) | Ratio of means |
| Agent systems | 1,000 | 168.27\,{\color[rgb]{0.5,0.5,0.5}\scriptstyle\pm}\,{\color[rgb]{0.5,0.5,0.5}8.24} | 55.47\,{\color[rgb]{0.5,0.5,0.5}\scriptstyle\pm}\,{\color[rgb]{0.5,0.5,0.5}2.96} | 3.03 |
| RAG | 1,000 | 157.46\,{\color[rgb]{0.5,0.5,0.5}\scriptstyle\pm}\,{\color[rgb]{0.5,0.5,0.5}3.73} | 45.72\,{\color[rgb]{0.5,0.5,0.5}\scriptstyle\pm}\,{\color[rgb]{0.5,0.5,0.5}2.12} | 3.44 |
| Reasoning | 1,000 | 151.37\,{\color[rgb]{0.5,0.5,0.5}\scriptstyle\pm}\,{\color[rgb]{0.5,0.5,0.5}3.31} | 42.07\,{\color[rgb]{0.5,0.5,0.5}\scriptstyle\pm}\,{\color[rgb]{0.5,0.5,0.5}0.73} | 3.60 |
| Multimodal learning | 1,000 | 151.04\,{\color[rgb]{0.5,0.5,0.5}\scriptstyle\pm}\,{\color[rgb]{0.5,0.5,0.5}1.48} | 42.34\,{\color[rgb]{0.5,0.5,0.5}\scriptstyle\pm}\,{\color[rgb]{0.5,0.5,0.5}3.61} | 3.57 |
| Safety and alignment | 1,000 | 152.02\,{\color[rgb]{0.5,0.5,0.5}\scriptstyle\pm}\,{\color[rgb]{0.5,0.5,0.5}1.15} | 41.77\,{\color[rgb]{0.5,0.5,0.5}\scriptstyle\pm}\,{\color[rgb]{0.5,0.5,0.5}1.23} | 3.64 |
| Efficient LLMs | 1,000 | 155.02\,{\color[rgb]{0.5,0.5,0.5}\scriptstyle\pm}\,{\color[rgb]{0.5,0.5,0.5}1.68} | 47.24\,{\color[rgb]{0.5,0.5,0.5}\scriptstyle\pm}\,{\color[rgb]{0.5,0.5,0.5}4.55} | 3.28 |
| Family | Reading directions | Papers |
| Reasoning / learning | Reasoning; reinforcement learning (RL) for reasoning; distillation; alignment | 19,507 |
| Agents / retrieval | Agent harnesses; self-improvement; multi-agent; RAG | 12,525 |
| Model construction | Pretraining; efficiency; code; multimodal | 22,440 |
| Data / assessment | Data synthesis; evaluation; safety; hallucination | 22,473 |
| Distinct union | All sixteen directions | 58,094 |
| Representation / input | Background | Method | Result | All facets |
| BM25: title | 30.88 | 24.40 | 24.84 | 26.62\,{\color[rgb]{0.52,0.52,0.52}\scriptstyle\pm\,6.48} |
| BM25: title + abstract | 58.21 | 32.14 | 42.69 | 44.07\,{\color[rgb]{0.52,0.52,0.52}\scriptstyle\pm\,6.65} |
| SPECTER2: full abstract | 61.19 | 40.26 | 54.43 | 51.78\,{\color[rgb]{0.52,0.52,0.52}\scriptstyle\pm\,6.76} |
| Qwen3-Embedding-4B | 64.89 | 46.24 | 57.20 | \mathbf{55.94}\,{\color[rgb]{0.52,0.52,0.52}\scriptstyle\pm\,6.66} |
| Qwen3-Embedding-8B | 63.84 | 43.56 | 59.21 | 55.37\,{\color[rgb]{0.52,0.52,0.52}\scriptstyle\pm\,6.83} |
| Readout | SPECTER2 | Qwen3-4B | Qwen3-8B | |||
| MRR | R@50 | MRR | R@50 | MRR | R@50 | |
| Frozen title | ||||||
| Frozen abstract | ||||||
| Title MLP | 7.92\,{\color[rgb]{0.52,0.52,0.52}\scriptstyle\pm\,0.03} | 25.88\,{\color[rgb]{0.52,0.52,0.52}\scriptstyle\pm\,0.20} | 9.41\,{\color[rgb]{0.52,0.52,0.52}\scriptstyle\pm\,0.10} | 29.08\,{\color[rgb]{0.52,0.52,0.52}\scriptstyle\pm\,0.45} | 9.50\,{\color[rgb]{0.52,0.52,0.52}\scriptstyle\pm\,0.11} | 29.83\,{\color[rgb]{0.52,0.52,0.52}\scriptstyle\pm\,0.03} |
| Abstract MLP | 13.72\,{\color[rgb]{0.52,0.52,0.52}\scriptstyle\pm\,0.08} | 37.35\,{\color[rgb]{0.52,0.52,0.52}\scriptstyle\pm\,0.17} | 15.35\,{\color[rgb]{0.52,0.52,0.52}\scriptstyle\pm\,0.13} | 39.21\,{\color[rgb]{0.52,0.52,0.52}\scriptstyle\pm\,0.33} | 15.35\,{\color[rgb]{0.52,0.52,0.52}\scriptstyle\pm\,0.21} | 39.17\,{\color[rgb]{0.52,0.52,0.52}\scriptstyle\pm\,0.43} |
| Mean GNN (1) | 13.81\,{\color[rgb]{0.52,0.52,0.52}\scriptstyle\pm\,0.11} † | 38.12\,{\color[rgb]{0.52,0.52,0.52}\scriptstyle\pm\,0.14} † | 15.59\,{\color[rgb]{0.52,0.52,0.52}\scriptstyle\pm\,0.12} | 40.48\,{\color[rgb]{0.52,0.52,0.52}\scriptstyle\pm\,0.36} | 15.70\,{\color[rgb]{0.52,0.52,0.52}\scriptstyle\pm\,0.30} | 40.58\,{\color[rgb]{0.52,0.52,0.52}\scriptstyle\pm\,0.12} |
| System | Primary task | Role relative to KnowMe |
| I. Personal literature discovery | ||
| Scholar Inbox [ Flicke et al., 2025 ] | Recommendations, feedback, maps | Personalized discovery comparison |
| PaperFlow [ Wang et al., 2026d ] | Daily profiles and adaptation | Sequential recommendation comparison |
| Westlake Scholar [ Pan et al., 2026 ] | Institutional scholarly workflows | Repository-centered discovery |
| II. Grounded scientific workflows | ||
| InnoEval [ Qiao et al., 2026b ] | Research-idea assessment | Evidence-intensive evaluation |
| Evidence selection | Schema | Valid IDs (%) | Quotes (%) | Specificity | Feasibility | |
| 20 | Full context | 14/16 | 100.0 | 94.8 | \mathbf{3.15}\,{\color[rgb]{0.52,0.52,0.52}\scriptstyle\pm\,0.19} | \mathbf{2.44}\,{\color[rgb]{0.52,0.52,0.52}\scriptstyle\pm\,0.24} |
| 20 | Lexical + recency | 13/16 | 100.0 | 98.5 | 3.13\,{\color[rgb]{0.52,0.52,0.52}\scriptstyle\pm\,0.23} | 2.42\,{\color[rgb]{0.52,0.52,0.52}\scriptstyle\pm\,0.32} |
| 50 | Full context | 14/16 | 100.0 | 95.6 | \mathbf{3.11}\,{\color[rgb]{0.52,0.52,0.52}\scriptstyle\pm\,0.21} | 2.43\,{\color[rgb]{0.52,0.52,0.52}\scriptstyle\pm\,0.23} |
| 50 | Lexical + recency | 15/16 | 100.0 | 97.4 | 3.04\,{\color[rgb]{0.52,0.52,0.52}\scriptstyle\pm\,0.27} | \mathbf{2.46}\,{\color[rgb]{0.52,0.52,0.52}\scriptstyle\pm\,0.28} |
Appendix figures & tables44 assets
Supplementary material from the paper’s appendix.
Appendix
| Host | Papers | Base (s) | Personal (s) | Personal RSS (GiB) |
| 24 GiB / 14 cores | 10,000 | 1.72\,{\color[rgb]{0.52,0.52,0.52}\scriptstyle\pm\,0.55} | 1.52\,{\color[rgb]{0.52,0.52,0.52}\scriptstyle\pm\,0.41} | \mathbf{1.02}\,{\color[rgb]{0.52,0.52,0.52}\scriptstyle\pm\,0.22} |
| 24 GiB / 14 cores | 50,000 | 10.76\,{\color[rgb]{0.52,0.52,0.52}\scriptstyle\pm\,0.29} | 12.08\,{\color[rgb]{0.52,0.52,0.52}\scriptstyle\pm\,0.61} | \mathbf{0.91}\,{\color[rgb]{0.52,0.52,0.52}\scriptstyle\pm\,0.08} |
| 24 GiB / 14 cores | 100,000 | 20.68\,{\color[rgb]{0.52,0.52,0.52}\scriptstyle\pm\,0.50} | 22.71\,{\color[rgb]{0.52,0.52,0.52}\scriptstyle\pm\,1.31} | \mathbf{1.02}\,{\color[rgb]{0.52,0.52,0.52}\scriptstyle\pm\,0.09} |
| 48 GiB / 12 cores | 10,000 | \mathbf{1.63}\,{\color[rgb]{0.52,0.52,0.52}\scriptstyle\pm\,0.58} | \mathbf{1.39}\,{\color[rgb]{0.52,0.52,0.52}\scriptstyle\pm\,0.38} | 1.45\,{\color[rgb]{0.52,0.52,0.52}\scriptstyle\pm\,0.32} |
| 48 GiB / 12 cores | 50,000 | \mathbf{9.82}\,{\color[rgb]{0.52,0.52,0.52}\scriptstyle\pm\,0.44} | \mathbf{11.12}\,{\color[rgb]{0.52,0.52,0.52}\scriptstyle\pm\,0.42} | 1.25\,{\color[rgb]{0.52,0.52,0.52}\scriptstyle\pm\,0.32} |
| 48 GiB / 12 cores | 100,000 | \mathbf{20.64}\,{\color[rgb]{0.52,0.52,0.52}\scriptstyle\pm\,0.68} | \mathbf{21.68}\,{\color[rgb]{0.52,0.52,0.52}\scriptstyle\pm\,0.11} | 1.09\,{\color[rgb]{0.52,0.52,0.52}\scriptstyle\pm\,0.02} |
| Direction | Actual records per request | Workers | QPS (requests/s) | Batch p95 latency (ms) |
| RAG | 500 | 1 | 28.46\,{\color[rgb]{0.5,0.5,0.5}\scriptstyle\pm}\,{\color[rgb]{0.5,0.5,0.5}0.21} | 637.74\,{\color[rgb]{0.5,0.5,0.5}\scriptstyle\pm}\,{\color[rgb]{0.5,0.5,0.5}28.47} |
| RAG | 500 | 4 | 103.19\,{\color[rgb]{0.5,0.5,0.5}\scriptstyle\pm}\,{\color[rgb]{0.5,0.5,0.5}2.71} | 228.07\,{\color[rgb]{0.5,0.5,0.5}\scriptstyle\pm}\,{\color[rgb]{0.5,0.5,0.5}6.51} |
| RAG | 500 | 8 | \mathbf{184.56}\,{\color[rgb]{0.5,0.5,0.5}\scriptstyle\pm}\,{\color[rgb]{0.5,0.5,0.5}8.90} | 113.02\,{\color[rgb]{0.5,0.5,0.5}\scriptstyle\pm}\,{\color[rgb]{0.5,0.5,0.5}6.80} |
| RAG | 1,000 | 1 | 12.03\,{\color[rgb]{0.5,0.5,0.5}\scriptstyle\pm}\,{\color[rgb]{0.5,0.5,0.5}0.25} | 1624.14\,{\color[rgb]{0.5,0.5,0.5}\scriptstyle\pm}\,{\color[rgb]{0.5,0.5,0.5}19.78} |
| RAG | 1,000 | 4 | 56.93\,{\color[rgb]{0.5,0.5,0.5}\scriptstyle\pm}\,{\color[rgb]{0.5,0.5,0.5}0.54} | 414.79\,{\color[rgb]{0.5,0.5,0.5}\scriptstyle\pm}\,{\color[rgb]{0.5,0.5,0.5}3.99} |
| RAG | 1,000 | 8 | \mathbf{103.33}\,{\color[rgb]{0.5,0.5,0.5}\scriptstyle\pm}\,{\color[rgb]{0.5,0.5,0.5}10.18} | 213.65\,{\color[rgb]{0.5,0.5,0.5}\scriptstyle\pm}\,{\color[rgb]{0.5,0.5,0.5}14.72} |
| Direction | Papers | Authorship | Coauthor | Shared author | Cites | Bound | Nonempty |
| Agents | 500 | 3,646 | 182,302 | 27 | 44 | 203 | 2 |
| RAG | 500 | 2,383 | 7,977 | 13 | 110 | 266 | 4 |
| Reasoning | 500 | 3,883 | 187,959 | 27 | 13 | 207 | 1 |
| Medical AI | 314 | 1,802 | 8,673 | 17 | 116 | 160 | 2 |
| Multimodal | 500 | 3,248 | 16,045 | 23 | 54 | 253 | 2 |
| Safety | 500 | 2,919 | 16,355 | 14 | 119 | 256 | 4 |
| Comparison window | Pre-2023 rate (%) | Post-2022 rate (%) |
| Equal complete windows: 2020–2022 / 2023–2025 | 9.53\,{\color[rgb]{0.52,0.52,0.52}\scriptstyle\pm\,1.88} | 25.16\,{\color[rgb]{0.52,0.52,0.52}\scriptstyle\pm\,1.46} |
| Full history: 2010–2022 / 2023–October 2026 | 22.74\,{\color[rgb]{0.52,0.52,0.52}\scriptstyle\pm\,0.72} | 29.50\,{\color[rgb]{0.52,0.52,0.52}\scriptstyle\pm\,2.05} |
| Output | Information retained |
| nodes.jsonl | Typed IDs; paper version/date; author identity basis and resolution status |
| edges.jsonl | Subject, predicate, object, direction, and source/derivation evidence |
| triples.tsv | Subject–predicate–object only; detailed provenance remains in the edge records |
| citation-coverage.jsonl | Coverage state; source time; bound work; available reference count |
| manifest.json | Scope, time, input/code digests, relation counts and output hashes |
| Direction | Matching papers | Top-100 replaced | Top-500 replaced | Top-1,000 replaced |
| Reasoning | 9,603 | 29 | 91 | 159 |
| RL reasoning | 4,830 | 25 | 83 | 117 |
| Distillation | 3,105 | 22 | 83 | 96 |
| Alignment | 7,565 | 32 | 105 | 156 |
| Agent harnesses | 3,826 | 28 | 68 | 102 |
| Self-improvement | 3,849 | 28 | 72 | 107 |
| Quantity | Observed value |
| A. Retained real source responses | |
| Complete source windows / response pages | 40 / 142 |
| All observed source entries | 158,868 |
| Occurrences retained by the expanded paper scope | 153,797 |
| Unique eligible paper IDs (and unique ID/version pairs) | 109,285 |
| Repeated eligible ID/version occurrences | 44,512 |
| Variant | Top-100 overlap (%) | Catalog exposure (%) | Network exposure (%) | Institution HHI |
| A. Content and public evidence | ||||
| Topic + age | 67.2 | 1.2 | 8.8 | 0.054 |
| Base score | 90.7 | 10.8 | 13.8 | 0.039 |
| B. Adding direction-conditioned evidence | ||||
| + personal author | 99.0 | 9.7 | 11.3 | 0.040 |
| + personal institution | 90.3 | 10.8 | 13.8 | 0.041 |
| Rank | Source and observed role | Abstract evidence |
| 1 | UniGFM-RAG [ Zhao et al., 2026b ] : method contribution | “a unified end-to-end GraphRAG framework” |
| 77 | Visual Tokens on Demand [ Rahman et al., 2026 ] : related-method context | “retrieval-augmented visual reasoning” appears among existing approaches; the proposed module budgets visual tokens. |
| 156 | Indigenous Community Cosmovision [ Fernández-Sabido et al., 2026 ] : application component | “deployed in a conversational interface with retrieval-augmented generation (RAG)” |
| Institution | Active authors | Effective support | ||
| Alibaba / Qwen | 609 | 970 | 1071.6 | 5.476 |
| NVIDIA | 221 | 379 | 343.0 | 5.355 |
| Google (DeepMind) | 340 | 306 | 437.3 | 5.325 |
| CMU | 670 | 365 | 552.9 | 5.318 |
| MIT | 501 | 209 | 408.6 | 5.205 |
| Tencent | 391 | 598 | 727.1 | 4.976 |
| Context | Graph ready | Prior only | Max. | Max. | ||
| LLM union | 109 | 5 | 0.975 | 0.931 | 0.391 | 0.312 |
| Reasoning | 105 | 9 | 0.984 | 0.962 | 0.388 | 0.310 |
| RL reasoning | 97 | 17 | 0.989 | 0.966 | 0.384 | 0.308 |
| Distillation | 83 | 31 | 0.963 | 0.939 | 0.382 | 0.305 |
| Alignment | 103 | 11 | 0.990 | 0.943 | 0.388 | 0.311 |
| Agents | 93 | 21 | 0.990 | 0.957 | 0.386 | 0.309 |
| Comparison | min. | median | min. | Top 100 (%) | Top 500 (%) | |
| Restart only | 32 | 0.9997 | 0.9999 | 0.9926 | 97.0 | 97.8 |
| Shrinkage only | 32 | 0.9999 | 1.0000 | 0.9964 | 97.0 | 98.6 |
| Decay only | 32 | 0.9986 | 0.9996 | 0.9870 | 97.0 | 97.8 |
| Joint grid | 416 | 0.9982 | 0.9997 | 0.9834 | 93.0 | 96.8 |
| Direction | Title papers | + abstract papers | Added (%) | Title contexts | + abstract contexts | New contexts |
| Reasoning | 2,227 | 9,604 | 76.8 | 85 | 105 | 20 |
| RL reasoning | 1,071 | 4,832 | 77.8 | 68 | 97 | 29 |
| Distillation | 1,156 | 3,105 | 62.8 | 65 | 83 | 18 |
| Alignment | 1,312 | 7,567 | 82.7 | 67 | 103 | 36 |
| Agent harnesses | 763 | 3,828 | 80.1 | 71 | 93 | 22 |
| Self-improvement | 1,187 | 3,849 | 69.2 | 77 | 88 | 11 |
| Historical context | MRR | Recall@10 | Recall@50 | Non-seed MRR | |
| Title history | — | 4.81 | 7.08 | 14.39 | 4.82 |
| Title + abstract history | — | 8.29 | 11.69 | 23.34 | 8.32 |
| Coauthor mixture | 0.5 | 8.18 | 11.50 | 22.86 | 8.21 |
| Semantic-gated mixture | 0.5 | 8.18 | 11.51 | 22.80 | 8.21 |
| Shuffled-neighbor control | 0 | 8.29 | 11.69 | 23.34 | 8.32 |
| Representation / input | Background | Method | Result | All facets |
| BM25: title | 30.88 | 24.40 | 24.84 | 26.62\,{\color[rgb]{0.52,0.52,0.52}\scriptstyle\pm\,6.48} |
| BM25: title + abstract | 58.21 | 32.14 | 42.69 | 44.07\,{\color[rgb]{0.52,0.52,0.52}\scriptstyle\pm\,6.65} |
| BM25: query facet | 50.45 | 29.22 | 46.30 | 41.82\,{\color[rgb]{0.52,0.52,0.52}\scriptstyle\pm\,5.90} |
| SPECTER2: full abstract | 61.19 | 40.26 | 54.43 | 51.78\,{\color[rgb]{0.52,0.52,0.52}\scriptstyle\pm\,6.76} |
| SPECTER2: query facet | 58.66 | 40.36 | 50.33 | 49.61\,{\color[rgb]{0.52,0.52,0.52}\scriptstyle\pm\,6.48} |
| SPECTER2: both facets | 58.55 | 40.11 | 52.99 | 50.39\,{\color[rgb]{0.52,0.52,0.52}\scriptstyle\pm\,5.43} |
| Configuration | NDCG@20 | Precision@10 | Linear-gain NDCG@20% |
| BM25: title | 30.92 | 16.00 | 35.95 |
| BM25: title + abstract | 47.69 | 28.60 | 52.39 |
| BM25: query facet | 45.65 | 27.40 | 51.11 |
| SPECTER2: full abstract | 55.70 | 34.60 | 61.59 |
| SPECTER2: query facet | 54.26 | 31.80 | 61.36 |
| SPECTER2: both facets | 53.86 | 33.20 | 61.20 |
| Context | ||||
| Coauthor mixture | 8.765 | 8.825 | 8.779 | 8.839 |
| Semantic-gated mixture | 8.765 | 8.812 | 8.811 | 8.849 |
| Shuffled-neighbor control | 8.765 | 8.326 | 7.308 | 4.043 |
| Readout | MRR | R@10 | R@50 | R@100 | NDCG@50 |
| Frozen title | 8.34 | 12.23 | 21.67 | 27.69 | 10.53 |
| Frozen abstract | 13.88 | 19.69 | 32.53 | 40.42 | 17.02 |
| Title MLP | 2.52\,{\color[rgb]{0.52,0.52,0.52}\scriptstyle\pm\,0.28} | 4.31\,{\color[rgb]{0.52,0.52,0.52}\scriptstyle\pm\,0.39} | 12.14\,{\color[rgb]{0.52,0.52,0.52}\scriptstyle\pm\,1.22} | 18.00\,{\color[rgb]{0.52,0.52,0.52}\scriptstyle\pm\,1.64} | 4.10\,{\color[rgb]{0.52,0.52,0.52}\scriptstyle\pm\,0.43} |
| Abstract MLP | 5.39\,{\color[rgb]{0.52,0.52,0.52}\scriptstyle\pm\,0.29} | 9.42\,{\color[rgb]{0.52,0.52,0.52}\scriptstyle\pm\,0.41} | 21.93\,{\color[rgb]{0.52,0.52,0.52}\scriptstyle\pm\,1.45} | 30.37\,{\color[rgb]{0.52,0.52,0.52}\scriptstyle\pm\,2.03} | 8.26\,{\color[rgb]{0.52,0.52,0.52}\scriptstyle\pm\,0.48} |
| Mean GNN (1) | 5.65\,{\color[rgb]{0.52,0.52,0.52}\scriptstyle\pm\,0.29} | 9.80\,{\color[rgb]{0.52,0.52,0.52}\scriptstyle\pm\,0.54} | 23.07\,{\color[rgb]{0.52,0.52,0.52}\scriptstyle\pm\,1.04} | 31.58\,{\color[rgb]{0.52,0.52,0.52}\scriptstyle\pm\,1.39} | 8.67\,{\color[rgb]{0.52,0.52,0.52}\scriptstyle\pm\,0.38} |
| Mean GNN (2) | 5.60\,{\color[rgb]{0.52,0.52,0.52}\scriptstyle\pm\,0.18} | 9.98\,{\color[rgb]{0.52,0.52,0.52}\scriptstyle\pm\,0.56} | 23.16\,{\color[rgb]{0.52,0.52,0.52}\scriptstyle\pm\,1.12} | 31.48\,{\color[rgb]{0.52,0.52,0.52}\scriptstyle\pm\,1.39} | 8.65\,{\color[rgb]{0.52,0.52,0.52}\scriptstyle\pm\,0.31} |
| Author group | Queries | Frozen abstract | Residual MLP | Residual GNN |
| Catalogue seeds | 118 | 8.82 | \mathbf{12.52}\,{\color[rgb]{0.52,0.52,0.52}\scriptstyle\pm\,0.58} | 11.67\,{\color[rgb]{0.52,0.52,0.52}\scriptstyle\pm\,0.60} |
| Outside catalogue | 15,576 | 15.94 | \mathbf{17.86}\,{\color[rgb]{0.52,0.52,0.52}\scriptstyle\pm\,0.15} | 17.85\,{\color[rgb]{0.52,0.52,0.52}\scriptstyle\pm\,0.07} |
| Supervised query authors | 10,484 | 12.15 | \mathbf{14.07}\,{\color[rgb]{0.52,0.52,0.52}\scriptstyle\pm\,0.10} | \mathbf{14.07}\,{\color[rgb]{0.52,0.52,0.52}\scriptstyle\pm\,0.07} |
| Other query authors | 5,210 | 23.42 | \mathbf{25.36}\,{\color[rgb]{0.52,0.52,0.52}\scriptstyle\pm\,0.26} | 25.33\,{\color[rgb]{0.52,0.52,0.52}\scriptstyle\pm\,0.09} |
| One context paper | 3,560 | 24.36 | \mathbf{26.31}\,{\color[rgb]{0.52,0.52,0.52}\scriptstyle\pm\,0.30} | 26.25\,{\color[rgb]{0.52,0.52,0.52}\scriptstyle\pm\,0.15} |
| Two context papers | 2,722 | 19.06 | 20.44\,{\color[rgb]{0.52,0.52,0.52}\scriptstyle\pm\,0.07} | \mathbf{20.57}\,{\color[rgb]{0.52,0.52,0.52}\scriptstyle\pm\,0.02} |
| Readout / input | NDCG@10 | NDCG@20 | P@10 | Linear NDCG at 20% |
| 27B relevance: title | 41.22\,{\color[rgb]{0.52,0.52,0.52}\scriptstyle\pm\,5.84} | 47.62 | 28.40 | 54.17 |
| 27B relevance: full | 59.80\,{\color[rgb]{0.52,0.52,0.52}\scriptstyle\pm\,6.11} | 63.92 | 42.80 | 68.88 |
| 27B relevance: query facet | \mathbf{61.51}\,{\color[rgb]{0.52,0.52,0.52}\scriptstyle\pm\,5.30} | 65.24 | 42.60 | 70.18 |
| 8B embedding: full | 55.37\,{\color[rgb]{0.52,0.52,0.52}\scriptstyle\pm\,6.83} | 60.06 | 38.60 | 65.10 |
| 8B embedding: query facet | 57.07\,{\color[rgb]{0.52,0.52,0.52}\scriptstyle\pm\,5.92} | 62.02 | 38.00 | 66.75 |
| BM25: title | 26.62\,{\color[rgb]{0.52,0.52,0.52}\scriptstyle\pm\,6.48} | 30.92 | 16.00 | 35.95 |
| Stratum | Primary | Secondary | Mention | None | Supported | |
| Title match | 640 | 365 | 109 | 15 | 151 | 472/640 |
| Abstract-only | 1280 | 400 | 356 | 98 | 426 | 753/1280 |
| No lexical match | 640 | 20 | 17 | 15 | 588 | 37/640 |
| Qwen3-32B input | NDCG@10 | NDCG@20 | P@10 | Linear NDCG at 20% |
| Title | 41.13\,{\color[rgb]{0.52,0.52,0.52}\scriptstyle\pm\,5.95} | 47.44 | 27.20 | 52.60 |
| Full | \mathbf{52.97}\,{\color[rgb]{0.52,0.52,0.52}\scriptstyle\pm\,6.27} | 58.26 | 37.00 | 63.55 |
| Query Facet | 52.25\,{\color[rgb]{0.52,0.52,0.52}\scriptstyle\pm\,5.25} | 56.95 | 36.40 | 61.90 |
| Input | NDCG@10 | NDCG@20 | P@10 | Linear NDCG at 20% |
| Title | -0.09\,{\color[rgb]{0.52,0.52,0.52}\scriptstyle\pm\,4.16} | -0.17\,{\color[rgb]{0.52,0.52,0.52}\scriptstyle\pm\,3.51} | -1.20\,{\color[rgb]{0.52,0.52,0.52}\scriptstyle\pm\,4.00} | -1.57\,{\color[rgb]{0.52,0.52,0.52}\scriptstyle\pm\,3.70} |
| Full | -6.84\,{\color[rgb]{0.52,0.52,0.52}\scriptstyle\pm\,3.64} | -5.66\,{\color[rgb]{0.52,0.52,0.52}\scriptstyle\pm\,3.65} | -5.80\,{\color[rgb]{0.52,0.52,0.52}\scriptstyle\pm\,2.98} | -5.32\,{\color[rgb]{0.52,0.52,0.52}\scriptstyle\pm\,3.03} |
| Query Facet | -9.26\,{\color[rgb]{0.52,0.52,0.52}\scriptstyle\pm\,3.36} | -8.29\,{\color[rgb]{0.52,0.52,0.52}\scriptstyle\pm\,3.29} | -6.20\,{\color[rgb]{0.52,0.52,0.52}\scriptstyle\pm\,3.03} | -8.28\,{\color[rgb]{0.52,0.52,0.52}\scriptstyle\pm\,2.99} |
| 27B 32B | Primary | Secondary | Mention | None |
| Primary | 675 | 47 | 8 | 55 |
| Secondary | 185 | 176 | 49 | 72 |
| Mention | 10 | 32 | 42 | 44 |
| None | 31 | 74 | 35 | 1025 |
| Contrast | Relevance Lower | Relevance Upper | Reference Lower | Reference Upper |
| Full context fixed-penalty baseline | ||||
| Full context fixed reference | ||||
| Add author catalogue | ||||
| Add topic-fit authors | ||||
| Add institution context | ||||
| Add network authors |
| Arm | Relevance | Reference cue | Substantive (%) | Available / |
| Jev: mean 95% half-width | ||||
| Fixed-penalty baseline | \mathbf{2.655}\,{\color[rgb]{0.52,0.52,0.52}\scriptstyle\pm\,0.038} | 1.881\,{\color[rgb]{0.52,0.52,0.52}\scriptstyle\pm\,0.013} | 93.8\,{\color[rgb]{0.52,0.52,0.52}\scriptstyle\pm\,1.8} | 800 / 800 |
| Author catalogue | 2.651\,{\color[rgb]{0.52,0.52,0.52}\scriptstyle\pm\,0.039} | 1.884\,{\color[rgb]{0.52,0.52,0.52}\scriptstyle\pm\,0.012} | \mathbf{93.9}\,{\color[rgb]{0.52,0.52,0.52}\scriptstyle\pm\,1.8} | 800 / 800 |
| Topic-fit authors | 2.646\,{\color[rgb]{0.52,0.52,0.52}\scriptstyle\pm\,0.038} | 1.884\,{\color[rgb]{0.52,0.52,0.52}\scriptstyle\pm\,0.012} | 93.6\,{\color[rgb]{0.52,0.52,0.52}\scriptstyle\pm\,1.8} | 800 / 800 |
| Institution context | 2.633\,{\color[rgb]{0.52,0.52,0.52}\scriptstyle\pm\,0.042} | 1.886\,{\color[rgb]{0.52,0.52,0.52}\scriptstyle\pm\,0.012} | 93.1\,{\color[rgb]{0.52,0.52,0.52}\scriptstyle\pm\,1.9} | 800 / 800 |
| Full context | 2.636\,{\color[rgb]{0.52,0.52,0.52}\scriptstyle\pm\,0.040} | \mathbf{1.887}\,{\color[rgb]{0.52,0.52,0.52}\scriptstyle\pm\,0.011} | 93.2\,{\color[rgb]{0.52,0.52,0.52}\scriptstyle\pm\,1.8} | 800 / 800 |