ExecTune: Effective Steering of Black-Box LLMs with Guide Models
A training method for small 'guide' models cuts inference costs up to 22% while matching larger models' accuracy on math and code — with no access to the big model's weights.

The Thesis
Most AI applications today call powerful, expensive models repeatedly through APIs — and those per-call costs add up fast, often exceeding what it would cost to train a smaller model outright. This paper introduces ExecTune, a way to train a small, cheap 'guide' model that tells a black-box large model (like Anthropic's Claude) exactly how to approach a problem, reducing wasted computation. The key insight is that what matters isn't just whether the guide's strategy is good — it's whether the big model can actually follow it, a property the authors call 'executability.' By optimizing directly for executability, ExecTune lets a cheaper model (Claude Haiku 3.5) outperform a more expensive one (Claude Sonnet 3.5) on math and coding benchmarks, and come within 1.7 percentage points of Claude Sonnet 4 at 38% lower cost. The catch: results are on standard benchmarks, and real-world gains depend heavily on how predictable the target black-box model's behavior is.
Catalyst
Black-box API pricing — where you pay per token at inference time — has become a significant operational cost for companies running AI in production at scale. At the same time, smaller open-weight models have become capable enough to act as effective orchestrators or planners, making the guide-core architecture newly practical. The recent proliferation of structured output features and function-calling APIs in models like Claude and GPT-4o also makes it easier to enforce the kind of structured strategies this approach depends on.
What's New
Earlier approaches to steering large models included simple prompting (just asking the model to think step-by-step), supervised fine-tuning on fixed datasets, and 'advisor' systems that prepend a hint to the prompt — none of which were trained with explicit awareness of whether the big model could actually execute the suggested strategy. The authors unify all these approaches under a single framework (Guide-Core Policies, or GCoP) and show they all fail to optimize the key variable: execution success. ExecTune adds a three-stage training loop — using outputs from a teacher model to filter good strategies, then fine-tuning on those, then applying reinforcement learning that rewards syntactic validity and successful task completion — directly targeting the gap between strategy quality and strategy executability.
The Counter
The benchmark improvements — up to 9.2% accuracy gain and 22.4% cost reduction — come from mathematical reasoning and code generation tasks, which are unusually well-structured and easily verifiable. Real enterprise workloads are messier: customer service, document analysis, and multi-step reasoning over proprietary data don't offer clean 'execution success' signals to optimize against. The executability metric is measured against specific Claude model versions; Anthropic updates its models frequently, and a guide trained on Haiku 3.5's behavior may degrade or need retraining after the next model revision, reintroducing the costs this approach claims to amortize. The paper also doesn't release training code or model weights, making independent replication difficult and raising questions about how sensitive results are to implementation choices. Finally, frontier models are getting cheaper rapidly — OpenAI and Anthropic have both cut prices significantly over the past year — which narrows the cost advantage window that makes this architecture compelling in the first place.
Longs
- AMZN — AWS hosts Anthropic's Claude API; cost-reduction techniques drive higher API volume
- NET (Cloudflare) — AI Gateway and Workers AI benefit from agentic, multi-model routing architectures
- SOUN (SoundHound AI) — edge AI orchestration for voice/task pipelines mirrors guide-core patterns
- PLTR — enterprise AI deployments where recurring inference cost is a procurement concern
- ANET (Arista Networks) — data center networking for high-throughput API traffic scales with agentic AI usage
Shorts
- Frontier model providers (Anthropic, OpenAI) at the margin — if guide-core systems let smaller, cheaper models replace larger ones for a significant share of tasks, average revenue per query falls
- Prompt engineering consultancies — ExecTune automates the strategy-design layer that human prompt engineers currently tune by hand
- Companies selling 'chain-of-thought' or 'reasoning' API add-ons — if a small guide model can replicate most of the benefit at lower cost, premium reasoning tiers lose pricing power
Enablers (Picks & Shovels)
- Anthropic's Claude API — the paper's primary benchmark environment; structured output and tool-use features are prerequisites
- Hugging Face Transformers and PEFT libraries — used for fine-tuning small guide models without full retraining
- OpenAI-compatible inference APIs — the guide-core abstraction generalizes to any API with predictable structured output behavior
- RLVR (Reinforcement Learning with Verifiable Rewards) tooling — the RL component of ExecTune relies on recent open-source RL-for-reasoning frameworks
Private Watchlist
- Anthropic (private) — Claude Haiku/Sonnet are the specific models benchmarked; cost-efficiency framing directly benefits their mid-tier SKUs
- Together AI (private) — inference platform where fine-tuned guide models could be hosted cheaply
- Cognition AI (private) — agentic coding systems would directly benefit from guide-core cost reduction
- Cohere (private) — enterprise API provider where inference cost reduction is a key selling point
Resources
The Paper
For large language models deployed through black-box APIs, recurring inference costs often exceed one-time training costs. This motivates composed agentic systems that amortize expensive reasoning into reusable intermediate representations. We study a broad class of such systems, termed Guide-Core Policies (GCoP), in which a guide model generates a structured strategy that is executed by a black-box core model. This abstraction subsumes base, supervised, and advisor-style approaches, which differ primarily in how the guide is trained. We formalize GCoP under a cost-sensitive utility objective and show that end-to-end performance is governed by guide-averaged executability: the probability that a strategy generated by the guide can be faithfully executed by the core. Our analysis shows that existing GCoP instantiations often fail to optimize executability under deployment constraints, resulting in brittle strategies and inefficient computation. Motivated by these insights, we propose ExecTune, a principled training recipe that combines teacher-guided acceptance sampling, supervised fine-tuning, and structure-aware reinforcement learning to directly optimize syntactic validity, execution success, and cost efficiency. Across mathematical reasoning and code-generation benchmarks, GCoP with ExecTune improves accuracy by up to 9.2% over prior state-of-the-art baselines while reducing inference cost by up to 22.4%. It enables Claude Haiku 3.5 to outperform Sonnet 3.5 on both math and code tasks, and to come within 1.7% absolute accuracy of Sonnet 4 at 38% lower cost. Beyond efficiency, GCoP also supports modular adaptation by updating the guide without retraining the core.