acceptodds
Withdrawn from ICLR 2027

Automated Algorithm Discovery through Mathematical Reformulation: A Study in Visual-Token Reduction

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Automated algorithm discovery typically searches within a fixed understanding of a task, mutating programs or recombining familiar heuristics. For algorithm-design tasks that admit useful mathematical abstraction, the mathematical formulation shapes algorithm design before implementation. Such a task may be expressed as different mathematical problems, each revealing different properties to preserve and different algorithmic structures. Each formulation admits multiple solution procedures with different states, updates, and outcomes. Our central insight is therefore to make both the mathematical formulation and its solution procedure explicit objects of search. We introduce a formulation-centric framework in which each proposal records a formulation–solution pair. An AI–Math–AI loop generates diverse formulations and a growing toolbox of mathematical mechanisms, retrieves relevant external knowledge, derives solution procedures, and implements them as executable programs. Its feedback mechanism balances exploitation and diversity: it allocates more search effort to mathematical areas associated with stronger proposals, while discouraging repeated exploration and redirecting attempts toward underexplored areas and tools. We study visual-token reduction in frozen multimodal large language models, where the framework discovers effective selectors grounded in reconstruction, information gain, and coverage. Their performance across tasks, token budgets, and model backbones demonstrates the practical value of exploring alternative mathematical views of the same problem. These results support searching over mathematical formulations and solution procedures as a productive direction for automated algorithm discovery.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Automated Algorithm Discovery, Token Reduce

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_S6h7OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes an AI–Math–AI framework that searches over mathematical formulations and solution procedures for automated algorithm discovery. Experiments on visual-token reduction demonstrate improvements over evolutionary baselines and existing token-selection methods across multiple benchmarks and model backbones.

Critical strengths
  1. The AI–Math–AI framework is well structured, connecting mathematical formulations, solution procedures, and executable algorithms in a clear workflow.

  2. The framework achieves better discovery scores than the evaluated evolutionary baselines, while the discovered algorithms consistently outperform existing visual-token reduction methods across multiple benchmarks and token budgets.

  3. The discovered algorithms demonstrate improvements across different model backbones and have interpretable mathematical mechanisms.

Critical weaknesses
  1. Its methodological novelty appears limited, as the main components largely build on existing techniques. The discovered algorithms also largely build on established mathematical methods, leaving the distinctive methodological contribution insufficiently demonstrated.

  2. The direct-mutation ablation removes multiple components simultaneously, making it difficult to isolate the contribution of mathematical formulation search, which is the paper's central methodological contribution.

  3. The search comparisons are matched by the number of attempts but not by total computational cost, and the evolutionary baselines lack comparable extended-budget evaluations. Furthermore, the reliance on single search runs makes it difficult to assess the stability of the reported improvements.

  4. The writing is fragmented, and the overall presentation would benefit from further improvement in clarity and coherence.

Questions for authors
  1. To isolate the contribution of mathematical formulation search, I encourage the authors to include a controlled ablation that removes only the formulation-search step while keeping all other components unchanged.

  2. The paper uses different search budgets across experiments, and the connection between the algorithm discovery results and the final downstream evaluation requires further clarification. The comparison with existing discovery methods is limited to 150 attempts, whereas the final evaluated algorithms were obtained from substantially longer searches. Could the authors clarify whether the proposed framework would still outperform competing discovery methods under comparable extended search budgets?

  3. How does the framework compare with the evolutionary baselines under matched total computational cost (e.g., tokens, API calls, or wall-clock time), including extended-budget runs for the baselines? Results averaged over multiple independent search runs, with variance reported, would also help establish whether the improvements are stable.

Reviewer_3LMHOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper uses an LLM to search for token-pruning algorithms under a fixed token budget. It starts with mathematical goals, such as coverage and reconstruction, and a toolbox of possible solution methods. Each attempt retrieves mathematical information, derives a selection rule, generates and checks code, and measures downstream benchmark accuracy. Search records guide later exploration and help avoid repeated ideas. The paper selects three algorithm families, FPS, ORGS, and Leverage, for detailed analysis and evaluation across benchmarks and models.

Critical strengths
  1. The reported pruning results appear competitive with the state of the art. The method produces concrete selectors that could be useful in practice.

  2. Choosing both a mathematical goal and a tool is an interesting way to explore algorithms beyond small changes to an existing program.

Critical weaknesses
  1. It is unclear how much initialization affects the results. The search starts with goals such as coverage, reconstruction, and information gain. These are reasonable choices, but they are also familiar ideas in token pruning. I could not find a clear test of how much the results depend on this starting set. Can the method still succeed if useful goals or tools are missing? This matters when judging how much the search discovers, rather than inherits from its setup.

  2. The search process and its results need a clearer explanation. The pruning results are stronger evidence than the current explanation of the discovery process. I would like to see which initial goals led to useful candidates and why the three final selectors were chosen.

  3. It is not clear how this approach would extend to other problems.

Questions for authors
  1. How sensitive are the results to the initial goals and toolbox? Do any ablations test removing useful goals or starting from a simple task description without domain-specific guidance?

  2. What is the benefit of using these preservation goals compared with searching directly for high accuracy (LM accuracy) under a token budget? How much do familiar pruning ideas limit the search?

  3. Could you summarize the attempts and best scores for each initial goal, and explain how the three final selectors were chosen? What was the best query-aware candidate from the full method? Were any candidates designed to meet several goals together?

  4. Is there evidence of successful search on another problem? If not, please clarify the scope of the claims. A short end-to-end example would also help me understand the method.

Reviewer_CPWcOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes an automated algorithm-discovery framework that searches over both mathematical formulations and solution procedures. Its AI–Math–AI pipeline generates formulations, retrieves mathematical knowledge, synthesizes executable algorithms, and uses evaluation feedback to guide exploration. Applied to visual-token pruning, it discovers three selectors and reports improvements over evolutionary search and existing pruning baselines.

Critical strengths
  1. The formulation-guided search approach is interesting and potentially useful.

  2. Strong results across multiple benchmarks, token budgets, and model backbones.

  3. Useful ablations examining the discovered selectors.

Critical weaknesses
  1. Novelty is a major concern. ORGS uses a residual-based selection mechanism closely related to GOTS, SPARE, and ResPrune [1–3]. The broader framework also overlaps conceptually with OptMaster [4]. None of these relevant works is discussed.

  2. The ablations do not isolate the contribution of mathematical formulation search. The direct-mutation baseline also removes tool selection, retrieval, and formulation–solution derivation.

  3. Search baselines use only one run each and match the number of attempts rather than computational cost. No variance across independent runs is reported.

  4. Algorithm discovery is demonstrated on only one problem (visual-token reduction), making the generality claim difficult to assess.

[1] Ling et al. (2026). GOTS: Greedy Orthogonal Token Selection for High-Resolution Vision-Language Models.

[2] Lee et al. (2026). Moving Beyond Diversity: Visual Token Pruning as Subspace Reconstruction for Efficient VLMs.

[3] Li et al. (2026). ResPrune: Text-Conditioned Subspace Reconstruction for Visual Token Pruning in Large Vision-Language Models.

[4] Lin et al. (2026). OptMaster: A DAG-Based Framework for Formulation and Heuristic Discovery in Optimization. ICML 2026.

Questions for authors
  1. How does ORGS differ from GOTS, SPARE, and ResPrune? Direct comparisons seem necessary.

  2. What is the specific methodological novelty over OptMaster?

  3. Can the authors isolate the contribution of formulation search and demonstrate results on another problem?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.