TaxoConf: Taxonomy-Guided Automatic Conference Program Organization

1 INTRODUCTION

• Each year, conferences must organize hundreds to thousands of accepted papers, and manual grouping, session naming, and scheduling consume considerable time and effort.SIGIR 2026 organizers arranged 657 accepted presentations into 51 oral sessions and 3 poster sessions while meeting hard constraints.Source: §1 · p. 1
• A deployable program must combine topical coherence with operational feasibility, although semantic grouping alone and constraint satisfaction alone each fall short.Constraints include session capacity, track compatibility, and presenter conflicts across parallel sessions.Source: §1 · p. 1
• TaxoConf separates semantic modeling from combinatorial scheduling, building a canonicalized multi-parent topic taxonomy and taxonomy-guided paper distance for constrained session assignment.Its workflow uses information-content-weighted taxonomy edges and optimal transport between paper frontier topics, then optimizes assignments subject to hard constraints.Source: §1 · pp. 1–2, Figure 1 · p. 2
• TaxoConf achieved the best coherence and human agreement while incurring zero constraint violations, and SIGIR 2026 organizers retained 91.8% of oral assignments and adopted the poster program without modification.The on-site survey of 47 attendees rated overall program quality 4.26 out of 5.Source: §1 · p. 2

2 PROBLEM FORMULATION

• Organizers are assumed to specify a session plan capable of accommodating the papers, with each session defined by its room, time slot, capacity bounds, and permitted tracks.Permitted tracks determine whether sessions may mix tracks or are restricted to one track.Source: §2 · p. 2
• The formulation distinguishes topical coherence from operational requirements such as presenter availability, session-capacity limits, and track compatibility.Paper titles and abstracts provide topical evidence, while presenter and track metadata inform operational constraints.Source: §2 · pp. 2–3
• The task is to assign every accepted paper to exactly one session, maximizing topical coherence over assignments that satisfy all mandatory operational requirements.The formulation expresses this as maximizing F_coh over the feasible assignment set Ω_op.Source: §2 · p. 3

3 CANONICALIZED MULTI-PARENT TAXONOMY CONSTRUCTION

• A faithful taxonomy must accommodate topics with multiple parents and reconcile equivalent topics discovered across branches and at different depths.Tree-based expansion cannot represent cross-cutting topics, while inconsistent granularity makes topic positions depend on branch structure rather than topical content.Source: §3 · p. 3
• TaxoConf iteratively proposes finer topics and resolves each proposed topic against a global registry, collapsing equivalent topics into shared nodes that can have multiple parents.Papers are routed to proposed subtopics as a multi-label classification, allowing a paper to follow several branches.Source: §3 · p. 3
• The LLM proposes child topics from the papers assigned to a node and the node’s canonical name and definition, which constrain proposals to refinements of that topic.Titles and abstracts provide corpus evidence for identifying present subtopics.Source: §3 · p. 3
• The global registry matches proposed topics across branches and recursion depths; equivalent mentions merge into existing nodes, while new canonical topics are added for expansion.Canonicalized children link to their current parent, and shared nodes accumulate multiple incoming edges.Source: §3 · p. 3
• Routing makes an independent binary decision for each child because child topics may overlap, allowing a paper to be assigned to zero, one, or multiple children.Routing uses the paper’s title and abstract together with the child topic’s name and definition.Source: §3 · p. 3
• Expansion stops when a node has fewer than the minimum paper count or reaches maximum depth; each paper’s frontier then removes redundant ancestors while retaining incomparable topics.The process ends when no queued node remains expandable under those conditions.Source: §3 · p. 4

4.1 Specificity-based Taxonomy Distance

• On a DAG, TaxoConf uses specificity-weighted shortest undirected paths and optimal transport between uniform frontier-topic distributions to measure paper distance.Edge length is the information-content gain from a broader topic to a narrower one; the shortest-path definition generalizes LCA distance to multi-parent taxonomies.Source: §4.1 · p. 4

4.2 Constraint-Aware Optimization Formulation

• The assignment model minimizes total within-session semantic distance, with session weights normalizing contributions by paper-pair count.Source: §4.2 · p. 4
• Hard constraints enforce complete assignments, capacity and activation rules, track and placement compatibility, and presenter-conflict separation across parallel sessions.Source: §4.2 · p. 5
• When feasible, branch-and-bound and branch-and-cut solvers return an optimal assignment; infeasible hard-constraint systems trigger a preference-weighted soft-constraint fallback.Source: §4.2 · p. 5

• An LLM generates three-to-eight-word session titles from paper information and taxonomy paths, capturing shared themes rather than concatenating keywords.Source: §5 · p. 5
• Time-slot alignment maximizes topical divergence among co-scheduled sessions using embeddings of session titles and paper titles and abstracts.Source: §5 · p. 5
• Within each time slot, sessions are ranked by predicted popularity from historical related-paper citations and paired in rank order with rooms ranked by capacity.Source: §5 · p. 5
• Poster placement uses a one-dimensional board order to balance presenter and topical proximity, with traversal order approximating physical proximity.Source: §5 · pp. 5–6

6.1 Experiment Setting

• The benchmarks use official human-curated oral sessions as reference partitions, while all methods receive titles and abstracts and share session inventories.Evaluation covers session coherence and outliers, agreement with human sessions, and capacity and presenter-conflict violations.Source: §6.1 · p. 6

6.2.1 Program Quality Evaluation.

• 4% on coherence, 12% on NMI, and 33% on cluster-F1 were average gains over the strongest baseline; TaxoConf ranked first on all five metrics across four benchmarks with zero constraint violations.The reported average NMI gains were stable across venue types, at 10%–14%.Source: Table 2 · p. 6, §6.2.1 · p. 6
• Optimization baselines were feasible but less aligned with human programs, whereas LLM baselines lacked reliable operational performance, including 96 and 58 violations for LLM-Direct and LLM-ReAct.LLM-ReAct was within about 4% of TaxoConf on coherence but trailed Embedding-ILP on NMI; the strongest optimization baseline used the same assignment program and constraints as TaxoConf.Source: §6.2.1 · pp. 6–7

6.2.2 Efficiency and Monetary Cost Comparison.

• 45.9 minutes and $6.30 were TaxoConf’s average runtime and cost, 34% and 38% lower than LLM-ReAct, respectively.Optimization baselines were fastest and had no API cost, while LLM-Direct cost $4.61 but produced lower-quality, constraint-violating programs.Source: §6.2.2 · p. 7, Figure 2 · p. 7

6.3 Deployment and Post-Launch Performance

• The SIGIR 2026 deployment covered 657 accepted papers, and its initial program had zero hard-constraint violations requiring correction.The output comprised 293 oral presentations in 51 sessions and 364 posters in three sessions; organizers reviewed, revised, and approved the initial program.Source: §6.3.1 · p. 7, §6.3 · p. 7

6.3.1 Post-Launch Performance.

• 91.8% of oral assignments were preserved, and the entire poster program required no post-editing; attendees rated overall program quality 4.26 out of 5.Only two of 51 oral sessions were regrouped; the lowest-rated aspect was room-size adequacy at 3.77, and the paper identifies popularity-based room allocation as a refinement target.Source: §6.3.1 · pp. 7–8, Figure 3 · p. 7

Conference Program Organization

• Human-in-the-loop systems use affinity judgments to support session construction, but demand additional input that limits scalability.Optimization methods enforce constraints but rely on flat similarities, while LLM-based approaches offer richer semantic organization without jointly enforcing conference-wide requirements [20].Source: §7 · p. 8
• TaxoConf couples taxonomy-derived session composition, titles, and paper distances with constraint-aware integer programming to target coherence and operational feasibility.Its approach is presented as avoiding the crowd effort required by human-in-the-loop systems.Source: §7 · p. 8

Taxonomy Construction

• Existing LLM taxonomy methods produce trees through branch-by-branch expansion, which cannot represent multi-parent topics and may duplicate equivalent topics across branches [21] [47] [51].TaxoConf instead canonicalizes proposed topics against a global registry, forming a conference-specific topic DAG with shared multi-parent nodes.Source: §7 · p. 8

8 CONCLUSION

• Across four conference benchmarks, TaxoConf produced the most coherent sessions and closest agreement with human-curated programs without constraint violations, and its SIGIR 2026 program was deployed after minor edits.The system builds a canonicalized multi-parent taxonomy, derives a taxonomy-aware paper distance, and uses integer programming for session assignment.Source: §8 · p. 8

A DETAILS OF THE EXACT AND FALLBACK SOLUTION

• Relaxed constraints use penalized nonnegative slack variables, while assignment completeness, presenter conflicts, and binary domains remain unrelaxed.The slack values identify the degree of soft-constraint violation; track and placement compatibility reduce assignment variables, and track decomposition yields independent subproblems.Source: A DETAILS OF THE EXACT AND FALLBACK… · p. 10
• The assignment problem is NP-hard with exponential worst-case branch-and-bound complexity, but practical instances are solved to optimality or a small certified optimality gap within a time limit.Compatibility pruning and independent per-track subproblems reduce the practical problem size.Source: A DETAILS OF THE EXACT AND FALLBACK… · p. 10

B DETAILS OF BOARD PLACEMENT FOR THE POSTER PROGRAM

• Poster boards are ordered by spectral seriation over taxonomy affinity, with optional must-link groups keeping posters by shared presenters or demonstration–companion relations together.Demonstrations are placed on powered boards; when no compatible powered board exists, the procedure reports the conflict or relaxes optional presenter proximity.Source: B DETAILS OF BOARD PLACEMENT FOR THE… · p. 10

C.1 Benchmark Construction

• The benchmarks use official conference programs as human-curated oral-session references, enriched with abstracts retrieved from OpenAlex; poster placement is evaluated in the deployment study instead.Session inventories and human paper-to-session groupings are established from the official programs.Source: §C.1 · p. 11

C.2 TaxoConf Configuration

• TaxoConf uses a bounded-depth taxonomy procedure, an exactly solved transportation LP for paper distance, and CP-SAT for assignment with an eight-worker, 300-second limit per instance.The soft-constraint fallback never triggers on these benchmarks; the configuration uses gpt-5.4-mini for LLM components.Source: §C.2 · p. 11

C.3 Baselines

• The ILP baselines share TaxoConf’s assignment formulation and solver budget but replace its taxonomy distance with LDA-topic or SPECTER2 cosine distance.LLM-Direct generates an assignment in one pass, whereas LLM-ReAct runs up to three propose–check–revise rounds; both use the same LLM backbone as TaxoConf.Source: §C.3 · p. 11

C.4 Evaluation

• Three model families provide LLM coherence scores, which are averaged across judges.The judges are gpt-5.5, claude-opus-4.8, and gemini-3.1-pro.Source: §C.4 · p. 11
• Caching each judge call by its full request content makes the evaluation deterministic and rerunnable.Source: §C.4 · p. 11
• Table 4 reports ablation results averaged over four benchmarks, with arrows marking the preferred direction and bold marking the best result.Source: Table 4 · p. 11

D ABLATION STUDY

• Replacing paper frontiers with all assigned topics causes the largest degradation: NMI falls from 0.761 to 0.692 and outliers rise from 12.8 to 25.6.Broad ancestors dilute the most specific topics.Source: D ABLATION STUDY · p. 11
• Removing multi-parent routing costs 0.016 NMI and disabling canonicalization costs 0.029; even the weakest variant beats Embedding-ILP on average NMI (0.692 vs. 0.681) and coherence (4.41 vs. 4.14).The taxonomy ablations are milder but consistent, and the full margin accumulates across all five components.Source: D ABLATION STUDY · pp. 11–12

E ATTENDEE SURVEY AT SIGIR 2026

• The 47 respondents included attendees, oral and poster presenters, organizers or volunteers, and participants across four career stages.The survey used QR codes and in-person recruitment, collected ratings on ten program aspects using a 5-point scale, and allowed free-text comments.Source: E ATTENDEE SURVEY AT SIGIR 2026 · p. 12
• Overall program quality averaged 4.26 out of 5, and every surveyed aspect had a median of 4.Ease of finding relevant sessions also averaged 4.26; room-size adequacy was lowest at 3.77, and 12 respondents mentioned crowded rooms for popular sessions.Source: E ATTENDEE SURVEY AT SIGIR 2026 · p. 12
• Respondents also suggested website improvements, including saved-session bookmarks, more visible mobile filters, and calendar export.These suggestions concern the online program website rather than session assignment.Source: E ATTENDEE SURVEY AT SIGIR 2026 · p. 12