AI4AI at Test-Time: Strong-to-Weak Capability Transfer via Harnesses

Summary (Overview)

• Core contribution: This paper introduces and systematically studies "strong-to-weak scaffolding"—a test-time capability transfer paradigm where a strong "builder" model constructs inference-time harnesses (scaffolds) that improve a weaker "target" model's performance without any parameter updates to the target.
• Key finding: Scaffolding nearly doubles the average target model performance from 0.49 to 0.91 macro-average accuracy on Theory-of-Mind benchmarks, with 100% of runs exceeding the baseline and all 11 builder configurations producing positive uplift.
• Mechanism: Gains come primarily from offloading unstable model reasoning into deterministic code, benchmark-specific routing, and strict answer-format enforcement—not from encouraging longer reasoning or broader sampling.
• Key principles discovered: Builder capability and reasoning effort dominate platform choice; target model benefits scale with its "headroom" (available correctable errors); validation efficiency is high (median 5 evaluations, minimal overfitting).
• Implications: This reframes capability transfer as a complement to training-time distillation, suggesting models can transfer cognitive structure through external harnesses rather than weight updates.

Introduction and Theoretical Foundation

The paper addresses a fundamental question in model deployment: When a small model fails on a task, is the failure due to insufficient internal capability or excessive cognitive load imposed by task presentation?

Drawing on cognitive load theory (Sweller, 1988), the authors identify two complementary routes to improving performance:

1. Make the model more capable — the dominant approach in distillation research (Hinton et al., 2015; Hsieh et al., 2023; Agarwal et al., 2024)
2. Make the task easier to solve — realized through inference-time harnesses: external structures with routing logic, prompt templates, verification checks, memory, and tool use

The paper formalizes strong-to-weak scaffolding as a distinct setting where a strong builder model $M_{build}$ constructs a scaffold for a fixed weaker target model $M_{tar}$, with the builder never seeing the hidden test set. This isolates a practical form of transfer where capability flows through the harness rather than model weights.

The Theory-of-Mind (ToM) domain is chosen as a stress test because it requires tracking nested beliefs, perspective shifts, hidden information, and Bayesian goal inference—demanding tasks for smaller models that nonetheless contain exploitable structure.

Methodology

Formal Setting

The builder searches over a space of possible scaffolds $\mathcal{S}$:

$$S^{\star} = \arg \max_{S \in \mathcal{S}} \operatorname{Acc}\bigl(S, M_{\mathrm{tar}}; \mathcal{T}\bigr)$$

Since the test set $\mathcal{T}$ is hidden, the builder uses validation performance as a proxy:

$$\hat{S} = \arg \max_{S \in \mathcal{S}_{\text{build}}} \operatorname{Acc}(S, M_{\mathrm{tar}}; \mathcal{V})$$

Experimental Setup

• Benchmarks: Four ToM datasets aggregated into a 3,900-item hidden test set:

  • BigToM (1,200 items): Binary belief/goal/action questions on observation status
  • Hi-ToM (1,200 items): Nested belief questions with recursion order 0–4, deception, multi-room tracking
  • MMToM-QA (600 items): Binary Bayesian goal/belief inference from action traces
  • MuMA-ToM (900 items): 3-choice multi-agent belief/social-goal/belief-of-goal questions

• Validation: Each builder receives a 195-item (5%) validation sample; test set is held out

• Models: 11 builder configurations (Opus-4.7 with 4 reasoning effort tiers, Sonnet-4.6, GPT-5.5, GPT-5.4-mini, Codex-5.3, Gemini-3.1-Pro, Gemini-3.5-flash, Grok-0.1); 2 target models (GPT-5.4-mini, Gemini-3.5-flash)

• Platforms: 3 agentic coding harnesses (Cursor, Claude Code, GPT Codex)

• Total: 72 experiment runs, each repeated 3 times for stability analysis

Baselines

• Vanilla: Direct model call without scaffolding (GPT-5.4-mini: 0.488; Gemini-3.5-flash: 0.761)
• Human-Inspired Harness: UserHarness framework (GPT-5.4-mini: 0.939; Gemini-3.5-flash: 0.941)

Empirical Validation / Results

Main Results (Aspect 0)

The best scaffold outperforms both vanilla baselines on all four benchmarks and matches UserHarness on BigToM (1.00 vs. 0.95), with remaining gaps on Hi-ToM (0.80 vs. 0.87), MMToM-QA (0.84 vs. 0.98), and MuMA-ToM (0.88 vs. 0.96).

Key Findings by Aspect

Cognitive-Load Reduction

Determinism fraction (share of items answered entirely by code/rules) strongly predicts accuracy:

$$r = 0.72 \text{ (Pearson correlation between determinism fraction and final accuracy)}$$

Benchmark ofloadability varies sharply: BigToM ≈ 0.94, Hi-ToM ≈ 0.51, MMToM-QA ≈ 0.44, MuMA-ToM ≈ 0.36.

Residual Errors

Top scaffolds fix 83% of baseline errors while breaking only 7% of correct items. Residual errors concentrate in:

• Hi-ToM: Accuracy declines from 0.999 (order 0) to 0.700 (order 4); deception reduces performance
• MMToM-QA: Bayesian goal-inference subtypes (qtype 2.1: 0.680)
• MuMA-ToM: Social-goal (0.872) and belief-of-goal (0.880) labels

Theoretical and Practical Implications

Theoretical Contributions

1. Formalizes strong-to-weak scaffolding as a distinct inference-time capability transfer setting, complementing training-time distillation
2. Identifies the "headroom law": Scaffolding acts as competence recovery—its payoff is governed by the target's available correctable errors, not model identity
3. Establishes cognitive-load reduction as the mechanistic core: Builders act as "compilers of task competence," spending one-time reasoning to encode task structure into executable form

Practical Design Principles

• Use the strongest available builder with high reasoning effort
• Spend only a modest number of validation evaluations (median 5 suffices)
• Prioritize cognitive offloading for provable sub-tasks
• Build several independent scaffolds and select/ensemble for complementary repairs
• Apply scaffolding selectively at the sub-task level for already-strong targets, gated by measured headroom

Evaluation Reframing

The setting reframes evaluation from "How well can a model solve a task?" to "How well can a stronger model construct the conditions under which a weaker model can solve it?"—creating a potential benchmark for builder-model capability that measures externalization of reasoning into procedures, tools, and constraints.

Conclusion

Strong-to-weak scaffolding works because a capable builder can act as a compiler of task competence: it spends a one-time reasoning budget to identify and encode task structure into an inference-time scaffold, after which a weaker and cheaper target model can execute the task at a level approaching much stronger models.

Key takeaways:

1. Scaffolding does not replace raw reasoning capability—it reallocates it: the builder performs structural reasoning once; the target handles residual model-dependent cases
2. The approach is strongest for structured sub-problems with explicit decision procedures, but weaker for genuinely hard reasoning cores (nested belief tracking, deception, counterfactual goal inference)
3. Two complementary routes to stronger systems exist: improving internal capability (training) and making tasks easier to execute (harness design)—future progress depends on both, and on the co-evolution of models and inference environments

Future Directions

• Testing on broader benchmark families with different mixtures of symbolic structure, ambiguity, and open-ended reasoning
• Developing strong-to-weak scaffolding as a standard benchmark for builder models with secondary metrics (validation usage, inference cost, scaffold complexity, robustness)
• Studying harness self-evolution: agents automatically improving the environment that structures their own reasoning
• Exploring the interaction between training-time and test-time capability transfer