# arXiv cs.AI: Ballad: Bandit-Based LLM Routing for Automated Heuristic Discovery [Canonical reading](https://escapemeridian.com/readings/arxiv-cs-ai-ballad-bandit-based-llm-routing-for-automated-79ee09ba) In their arXiv abstract, Samidha Verma, Ankit Anand and Sayan Ranu argue that larger, costlier language models do not consistently produce better heuristics for automated heuristic discovery. They introduce Ballad, an online bandit algorithm that adaptively selects among models as a search proceeds. Because one model’s output can shape later calls, the authors say rewards change and interact across the evolving set of heuristics, complicating model selection. Ballad combines a policy favoring rare, high-scoring heuristics with credit propagation through their DAG-based genealogy and discounting of stale observations. Across five NP-hard combinatorial optimization problems and four models, the authors report that Ballad matched the best individual model in hindsight and sometimes surpassed every individual model. They also report API cost reductions of up to 65.6% compared with always using the strongest model. Escape Meridian’s interpretation is that these results suggest model choice during a search may matter alongside model capability. The abstract does not establish how broadly the reported performance generalizes. Headline and summary by Escape Meridian, prepared by Luna from the attributed source. A source check is not independent confirmation of the source’s claims. ## Original source arXiv cs.AI · Article Original source date: Nov 12, 2026 (2026-11-12). Date basis: Date basis not provided. Added to Escape Meridian: 2026-11-13T11:02:34.890Z. First observed: 2026-11-13T07:39:26.118Z. Retrieved: 2026-11-13T11:01:38.564Z. - [Ballad: Bandit-Based LLM Routing for Automated Heuristic Discovery](https://arxiv.org/abs/2611.17241) ## Limits The supplied evidence is the arXiv page and abstract, not the paper’s full methods or results; it does not provide details needed to assess the reported comparisons or their broader generalizability. ## Open question How does Ballad perform on problems and model pools beyond the five NP-hard combinatorial optimization problems and four models reported in the abstract? ## Source-stated The authors report finding no consistent advantage for larger, more expensive models in automated heuristic discovery across multiple NP-hard combinatorial optimization problems. > across multiple NP-hard combinatorial optimization problems, we find no consistent advantage for larger, costlier models. - [Supporting source](https://arxiv.org/abs/2611.17241) ## Source-stated The authors describe Ballad as an online bandit algorithm that adaptively orchestrates a pool of language models. > an online bandit algorithm that learns to orchestrate a pool of LLMs into an ensemble - [Supporting source](https://arxiv.org/abs/2611.17241) ## Source-stated The authors report that Ballad matched the best individual model in hindsight and sometimes surpassed every individual model across five problems and four models. > Across five NP-hard combinatorial optimization problems and a pool of four models of varying capability and cost, Ballad not only matches the best individual LLM in hindsight but, in several cases, surpasses every individual model - [Supporting source](https://arxiv.org/abs/2611.17241) ## Source-stated The authors report API cost reductions of up to 65.6% compared with always using the strongest model. > while reducing API cost by up to 65.6% relative to always using the strongest model. - [Supporting source](https://arxiv.org/abs/2611.17241) ## Escape Meridian interpretation The results suggest that adaptive model selection may matter during heuristic search, not just the capability of any single model. > The optimal LLM may shift over the course of a run - [Supporting source](https://arxiv.org/abs/2611.17241) ## Source and contributor context Verma, Samidha, Anand, Ankit, Ranu, Sayan · Article author These are credits stated by the source; they do not establish a profile match or authorship of every passage. ## Source media No suitable media was found by the supported page extraction. ## Terms and references Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement. ## Large language model A language model with many learned numerical settings, used to process and generate language. - [Definition reference](https://developers.google.com/machine-learning/glossary#large-language-model) ## API An application programming interface: a defined way for software to request data or actions from another system. - [Definition reference](https://developer.mozilla.org/en-US/docs/Glossary/API) ## arXiv arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed. - [Explore this context](/entities/arxiv) - [Reference](https://arxiv.org/) - [Reference](https://info.arxiv.org/help/api/index.html) ## Editorial provenance Version 1. Publication ID: 041ffda3-02b2-46ca-84a3-f5da8abfd69e. Prepared by gpt-6-luna; grounding review by gpt-6-luna. Source check: passed by anthropic/claude-opus-5.5 on 2026-11-13T11:02:34.890Z. This checks fidelity to the source, not whether its claims are independently true. Editorial selection: accept; reasons: material\_development. Observation: 00cf43d2e1f5c17226e20ae49d7dc8927437cb58d1b9998b6568e79397eedec7. Source text hash: ecd802b9e80c9b1ed8408c281149c310f333283b44a190e63085f5f5176e2fda. ## Revision and source-check history ### Version 1 — 2026-11-13T11:02:34.890Z In their arXiv abstract, Samidha Verma, Ankit Anand and Sayan Ranu argue that larger, costlier language models do not consistently produce better heuristics for automated heuristic discovery. They introduce Ballad, an online bandit algorithm that adaptively selects among models as a search proceeds. Because one model’s output can shape later calls, the authors say rewards change and interact across the evolving set of heuristics, complicating model selection. Ballad combines a policy favoring rare, high-scoring heuristics with credit propagation through their DAG-based genealogy and discounting of stale observations. Across five NP-hard combinatorial optimization problems and four models, the authors report that Ballad matched the best individual model in hindsight and sometimes surpassed every individual model. They also report API cost reductions of up to 65.6% compared with always using the strongest model. Escape Meridian’s interpretation is that these results suggest model choice during a search may matter alongside model capability. The abstract does not establish how broadly the reported performance generalizes. ## Original source arXiv cs.AI · Article Original source date: Nov 12, 2026 (2026-11-12). Date basis: Date basis not provided. Added to Escape Meridian: 2026-11-13T11:02:34.890Z. First observed: 2026-11-13T07:39:26.118Z. Retrieved: 2026-11-13T11:01:38.564Z. - [Ballad: Bandit-Based LLM Routing for Automated Heuristic Discovery](https://arxiv.org/abs/2611.17241) ## Limits The supplied evidence is the arXiv page and abstract, not the paper’s full methods or results; it does not provide details needed to assess the reported comparisons or their broader generalizability. ## Open question How does Ballad perform on problems and model pools beyond the five NP-hard combinatorial optimization problems and four models reported in the abstract? ## Source-stated The authors report finding no consistent advantage for larger, more expensive models in automated heuristic discovery across multiple NP-hard combinatorial optimization problems. > across multiple NP-hard combinatorial optimization problems, we find no consistent advantage for larger, costlier models. - [Supporting source](https://arxiv.org/abs/2611.17241) ## Source-stated The authors describe Ballad as an online bandit algorithm that adaptively orchestrates a pool of language models. > an online bandit algorithm that learns to orchestrate a pool of LLMs into an ensemble - [Supporting source](https://arxiv.org/abs/2611.17241) ## Source-stated The authors report that Ballad matched the best individual model in hindsight and sometimes surpassed every individual model across five problems and four models. > Across five NP-hard combinatorial optimization problems and a pool of four models of varying capability and cost, Ballad not only matches the best individual LLM in hindsight but, in several cases, surpasses every individual model - [Supporting source](https://arxiv.org/abs/2611.17241) ## Source-stated The authors report API cost reductions of up to 65.6% compared with always using the strongest model. > while reducing API cost by up to 65.6% relative to always using the strongest model. - [Supporting source](https://arxiv.org/abs/2611.17241) ## Escape Meridian interpretation The results suggest that adaptive model selection may matter during heuristic search, not just the capability of any single model. > The optimal LLM may shift over the course of a run - [Supporting source](https://arxiv.org/abs/2611.17241) ## Source and contributor context Verma, Samidha, Anand, Ankit, Ranu, Sayan · Article author These are credits stated by the source; they do not establish a profile match or authorship of every passage. ## Source media No suitable media was found by the supported page extraction. ## Terms and references Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement. ## Large language model A language model with many learned numerical settings, used to process and generate language. - [Definition reference](https://developers.google.com/machine-learning/glossary#large-language-model) ## API An application programming interface: a defined way for software to request data or actions from another system. - [Definition reference](https://developer.mozilla.org/en-US/docs/Glossary/API) ## arXiv arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed. - [Explore this context](/entities/arxiv) - [Reference](https://arxiv.org/) - [Reference](https://info.arxiv.org/help/api/index.html) ## Editorial provenance Version 1. Publication ID: 041ffda3-02b2-46ca-84a3-f5da8abfd69e. Prepared by gpt-6-luna; grounding review by gpt-6-luna. Source check: passed by anthropic/claude-opus-5.5 on 2026-11-13T11:02:34.890Z. This checks fidelity to the source, not whether its claims are independently true. Editorial selection: accept; reasons: material\_development. Observation: 00cf43d2e1f5c17226e20ae49d7dc8927437cb58d1b9998b6568e79397eedec7. Source text hash: ecd802b9e80c9b1ed8408c281149c310f333283b44a190e63085f5f5176e2fda. ### passed — 2026-11-13T11:02:34.890Z — anthropic/claude-opus-5.5 Publication: 041ffda3-02b2-46ca-84a3-f5da8abfd69e attribution — pass: The draft attributes the findings and Ballad to the listed authors, Samidha Verma, Ankit Anand and Sayan Ranu. It keeps reported results as the authors' claims, and the interpretation is labeled as Escape Meridian's. > Authors:Samidha Verma, Ankit Anand, Sayan Ranu summary — pass: The draft leads with the authors' empirical challenge to the assumption that bigger, costlier models are better. It explains Ballad's purpose (choosing among models as a search proceeds), the difficulty (one model's output shapes later calls), its three mechanisms, and the reported results. Technical terms are kept exactly for tooltips. The title and question are faithful, and the question raises a real open issue about generalization. > A natural assumption is that more powerful and more expensive models consistently yield better heuristics. We challenge this empirically date — pass: originalPublished 2026-11-12 matches the v1 submission date on the arXiv page. > \[Submitted on 12 Nov 2026\] support — pass: Each source-stated claim matches its excerpt: the no-consistent-advantage finding, the bandit description, matching or surpassing models across five problems and four models, and the up to 65.6% API cost reduction. The labeled interpretation is hedged with 'may' and grounded in the source's statement that the optimal LLM may shift during a run. The limitation about generalization is appropriate and does not overreach. > The optimal LLM may shift over the course of a run media — pass: The selected media is the verified direct arXiv abstract page for this paper. It is the most direct way to access the source. ## Recorded references Explicit citations in public readings and public file attachments to this reading. Shared topics are not citations. Other artifact types need an explicit reference adapter. No public references recorded here yet. [Reference data (JSON)](/readings/79ee09bafae9a45b8564225e1b719aac250350f7387fb1c8c289d87db7d18ffb/references.json) ## Corrections and challenges [Send a correction or an opposing source](https://escapemeridian.com/corrections/) [JSON record](https://escapemeridian.com/readings/arxiv-cs-ai-ballad-bandit-based-llm-routing-for-automated-79ee09ba.json)