TaxoConf: Taxonomy-Guided Automatic Conference Program Organization

Scientific Impact Assessment: TaxoConf

1. Core Contribution

TaxoConf addresses an under-automated step in the academic publication pipeline: assembling accepted papers into a session-based technical program. While submission, reviewing, and reviewer-assignment have received extensive automation attention, the downstream program-organization task has remained largely manual. The paper's central insight is to separate semantic modeling from combinatorial scheduling: it builds a conference-specific topic taxonomy, derives a principled paper-to-paper distance from it, and feeds that distance into a constraint-aware integer program.

Three technical contributions stand out: (1) a canonicalized multi-parent taxonomy construction method that produces a topic DAG (rather than a tree), collapsing equivalent topics discovered across branches into shared multi-parent nodes; (2) a specificity-weighted optimal-transport distance where edge lengths equal information-content gain and each paper is a distribution over frontier topics, compared via 1-Wasserstein distance; and (3) an ILP formulation encoding capacity, track, placement, and presenter-conflict constraints with a soft-constraint fallback. A refinement stage handles titles, time-slot divergence, room allocation, and poster board placement.

2. Methodological Rigor

The evaluation design is thoughtful. The optimization baselines (Topic-ILP, Embedding-ILP) solve the *identical* assignment program under identical constraints, differing only in the distance function — cleanly isolating the taxonomy's contribution. The ablation study (Appendix D) systematically disables each of five components, showing each matters, with frontier representation and optimal transport contributing most. Benchmarks are constructed from four real 2025 conferences with held-out human-curated sessions as reference, and precautions against leakage (shuffled identifiers, hidden session names) are documented.

Weaknesses: the LLM coherence metric is itself LLM-judged (mitigated by a three-model panel, but still circular relative to LLM-based baselines). No statistical significance testing or error bars appear in Table 2. Human-agreement metrics (NMI 0.76, ARI 0.24) treat the committee partition as ground truth, though the paper reasonably argues that coherent-but-different groupings can be legitimate. The deployment "survey" (N=47, convenience sample, self-selected) is suggestive rather than rigorous. The theoretical claims (metric validity of the Wasserstein distance, IC monotonicity along edges) are stated cleanly and are correct, though not deep.

3. Potential Impact

The most compelling evidence is the real deployment at SIGIR 2026: 91.8% of oral assignments and 100% of poster placements were preserved by organizers, with zero hard-constraint violations requiring correction. Additional slated adoptions (VLDB 2026, SIGIR 2027, ICDE 2027, DASFAA 2027, IJCNN 2026) signal genuine practitioner demand. A publicly deployed web application (taxoconf.com) with released datasets lowers the barrier to adoption substantially.

The direct impact is a useful, deployable tool for a recurring pan-disciplinary chore. The broader research impact is more modest but real: the canonicalized multi-parent taxonomy construction and the specificity-weighted OT distance are reusable primitives for scientific-document organization, literature-review structuring, and topic modeling beyond conference scheduling. The framing of "separate semantic modeling from combinatorial scheduling" is a clean design pattern others may adopt.

4. Timeliness & Relevance

The paper rides the wave of LLM-assisted document organization while addressing a concrete limitation of pure prompting approaches — their inability to enforce hard combinatorial constraints. This hybrid neuro-symbolic pattern (LLM for semantics, solver for constraints) is a timely and increasingly popular design. The task itself is perennial rather than trendy, which bodes well for sustained relevance but caps the "hot topic" ceiling.

5. Strengths & Limitations

Strengths:

Limitations:

Additional Observations

The paper is honest about failure modes (room sizing) and grounds its refinement claims in the deployment. Reproducibility is aided by released code/data, though the LLM backbone (gpt-5.4-mini) and judge models are proprietary and version-dependent, which may hamper exact replication over time. Resource requirements are modest (~46 min, ~$6.30 per conference on a single server), making the work highly accessible to extend. The contribution is best characterized as a strong, well-executed applied systems paper with a couple of reusable methodological nuggets — high practical impact within a bounded niche, moderate influence on core methods research.