Smart Matching: How to Assign Reviewers to Papers Automatically

Every conference chair knows the pain: 300 submissions, 80 reviewers, 3 reviewers per paper. That's 900 assignments to get right. And it's harder than it looks — reviewers have overlapping expertise, uneven workloads, conflicts of interest, and limited capacity. Assigning the best reviewer to one paper can leave another paper with no qualified options. The real problem isn't finding good matches for individual papers, it's optimizing across all papers and all reviewers simultaneously, subject to all those constraints at once.

How Smart Matching Works

Smart Matching solves this in three stages: score every possible reviewer-paper pair, filter out conflicts of interest, then find the globally optimal assignment.

Stage 1: Scoring

Every eligible reviewer is scored against every paper. The score (0-100) combines two signals:

Text similarity (up to 70 points) compares the paper's title and abstract against the reviewer's expertise profile. Two methods are available:

• BM25 — a proven text retrieval algorithm used by search engines and systems like OpenReview's TPMS. Free, runs in seconds.
• AI Semantic — uses OpenAI embeddings to understand that "NLP" and "natural language processing" mean the same thing, or that "deep learning" relates to "neural networks." Catches matches that pure text matching misses. Paper titles, abstracts, and reviewer tags are sent to OpenAI for embedding — use BM25 instead if data privacy is a concern.

Keyword overlap (up to 30 points) directly compares the paper's keywords against the reviewer's expertise tags — including system keywords, AI-discovered tags, and user-added tags. Exact phrase matches score highest, with stemmed word-level fallback for partial overlaps (e.g., "computing" matches "computational").

Stage 2: Conflict Filtering

Conflicts of interest are enforced before any assignment is made. Reviewers who are co-authors of a paper's authors (checked against the last 3 years of co-authorship data) or who have declared conflicts are excluded — the algorithm cannot override this.

Stage 3: Global Optimization

This is where Smart Matching differs from manual assignment. Instead of assigning one paper at a time, it solves for all assignments simultaneously using Min-Cost Max-Flow — a network optimization algorithm used by major conference systems like OpenReview (NeurIPS, ICML). The algorithm finds the assignment that maximizes total match quality across the entire conference, making the right trade-offs automatically when two papers compete for the same reviewer.

For typical conference sizes (50-500 papers, 20-200 reviewers), this completes in milliseconds.

Flexibility for Real Conferences

Academic conferences aren't clean optimization problems. Chairs need control over workload limits, whether to fill gaps or start fresh, and how to balance match quality against even workload distribution. Smart Matching supports all of these — see Auto Assign Reviewers for the full set of options.

The Best of Both Worlds

Smart Matching isn't meant to replace human judgment — it handles the 90% that's tedious so chairs can focus on the 10% that requires expertise. Let the algorithm generate assignments, review the suggestions, swap out any that don't feel right, then apply. Or make your key assignments by hand first, then run "Fill gaps" to handle the rest. The chair stays in control while the algorithm does the heavy lifting.

Before and After

---

Related Docs

• Auto Assign Reviewers — Step-by-step guide to running Smart Matching
• Smart Matching Algorithm — Technical deep-dive into scoring and optimization
• AI Research Profile Discovery — How reviewer expertise profiles are built
• Conflict of Interest Enforcement — COI rules and detection