# Evolving in Thought Space: Training a Small Model at Test Time Unlocks Better Discoveries (Guidance-TTT)

> Guidance-TTT decouples strategic adaptation from solution generation via a small trained guidance model, achieving state-of-the-art results across four scientific discovery domains at 58% lower cost.

- **Source:** [arXiv](https://arxiv.org/abs/2610.06269)
- **Published:** 2026-11-07
- **Permalink:** https://picx.dev/p/txYkn4
- **Whiteboard:** https://picx.dev/p/txYkn4/image

## Summary

# Summary of "Evolving in Thought Space: Training a Small Model at Test Time Unlocks Better Discoveries"

## Summary (Overview)

- **Key Contribution**: Introduces **Guidance-TTT**, a test-time reinforcement learning framework that decouples strategic adaptation from solution generation in LLM-based scientific discovery. A compact guidance model (8B or 14B parameters) proposes high-level changes, while a frozen execution model (GLM-5.2) implements them as complete solutions.

- **Core Innovation**: Training occurs in "thought space" — the guidance model learns to produce natural-language strategic modifications rather than complete programs, reducing adaptation cost and improving credit assignment by separating strategy quality from implementation quality.

- **Performance Results**: Outperforms prior work across all four domains tested: Polyomino Packing (91.89 score vs. previous SOTA 89.40), Lasso path optimization (0.1739 vs. 0.1243), AHC058 (850.08M points vs. 849.33M), and TriMul GPU kernel (1129 μs vs. 1131 μs), approaching the public rank-1 reference (1128 μs).

- **Cost Efficiency**: When using GPT-OSS-120B as executor, adaptation cost drops by ~58% compared to directly adapting the 120B model ($120 vs. $283), with comparable scores (83.19 vs. 83.72).

## Introduction and Theoretical Foundation

The paper addresses a fundamental tension in LLM-based scientific discovery systems. Prior approaches fall into two categories:

1. **Self-evolution methods** (FunSearch, AlphaEvolve, CORAL) keep models frozen and use verifier feedback only for selection and prompt conditioning
2. **Test-time training (TTT) methods** (TTT-Discover) update model parameters during inference to adapt generation policies

TTT-Discover's approach creates two critical problems:
- **Cost**: Reliable solution generation requires large models, making test-time adaptation expensive (memory, computation, optimizer states)
- **Credit assignment**: Outcome-level verifier feedback conflates high-level strategy quality with low-level implementation quality — a promising strategy fails if its implementation is faulty

The authors draw an analogy to human expertise: difficult problems are often solved through a small number of critical strategic decisions rather than reconsidering every implementation detail. This motivates separating strategic guidance from solution execution.

## Methodology

### Problem Setup
Following TTT-Discover, the framework considers a fixed discovery problem with:
- Task description $d$
- Executable verifier $V_d$ 
- Candidate solutions $C \in \mathcal{C}_d$ with reward $R = V_d(C)$

The goal is to find:
$$C^{\star} = \arg \max_{(C_i, R_i) \in \mathcal{A}_U} R_i$$

where $\mathcal{A}_U$ is the archive after $U$ updates.

### The Guidance-TTT Loop (Algorithm 1)

1. **Parent Selection**: Uses PUCT (Predictor + UCT) sampling:
$$Q(i) = \begin{cases} v(i), & \mathcal{C}(i) = \varnothing \\ \max_{j \in \mathcal{C}(i)} v(j), & \text{otherwise} \end{cases}$$

$$\text{PUCT}(i) = Q(i) + c s P(i) \frac{\sqrt{1 + N_{\text{exp}}}}{1 + n_i}$$

where $n_i$ is times node $i$ expanded, $N_{\text{exp}}$ is total expansions, $P(i)$ is rank-based prior, and $s$ normalizes exploration by score range.

2. **Guidance Generation**: The guidance model proposes $K$ high-level modifications conditioned on compact summaries of the parent solution.

3. **Execution**: The frozen execution model applies each modification to the exact parent solution.

4. **Verification**: Children candidates are evaluated by the verifier.

5. **Policy Update**: Only the guidance model is updated using group-relative RL with an entropic objective:

$$J_\beta(\theta \mid p) = \frac{1}{\beta} \log \mathbb{E}_{g \sim \pi_\theta(\cdot | s_p)} \mathbb{E}_{C \sim E(\cdot | C_p, g)} \left[ \exp(\beta V_d(C)) \right]$$

where $\pi_\theta$ is the guidance policy, $E$ is the frozen executor, and $\beta > 0$ controls emphasis on high rewards.

### Experimental Configuration
- Guidance model: Qwen3-8B (or 14B) with rank-32 LoRA
- Execution model: GLM-5.2 (frozen)
- 30 updates, $B = 8$ parents, $K = 16$ rollouts per parent
- No web access; seeds from one-shot execution-model calls

## Empirical Validation / Results

### Main Results (Table 2)

| Task | Previous Best | Guidance-TTT | Improvement |
|------|--------------|---------------|-------------|
| Polyomino | 89.40 (CORAL, web-enabled) | **91.89** | +2.49 |
| Lasso | 0.1243 (SimpleTES) | **0.1739** | +39.8% |
| AHC058 | 849,325,750 (SimpleTES) | **850,082,731** | +756,981 |
| TriMul | 1131 μs (K-Search) | **1129 μs** | -2 μs |

### Key Ablations (Table 3)

| Setting | Polyomino ↑ | TriMul latency (μs) ↓ |
|---------|-------------|----------------------|
| **Guidance-TTT** | **91.89** | **1225.58** |
| Frozen guidance | 84.85 | 1256.62 |
| Direct solution training | 41.63 | 9664.91 |
| Extra parent context (G-Top2) | 82.12 | 1166.82 |

Key findings:
- **Frozen guidance degradation**: Hierarchical decomposition alone insufficient; learning the guidance policy adds significant gains
- **Direct solution training failure**: Same 8B budget applied to complete solution generation performs dramatically worse (41.63 vs. 91.89), validating the benefit of learning in thought space
- **Context sensitivity**: Adding second-parent context helps TriMul (1166.82 μs) but hurts Polyomino (82.12), suggesting cross-parent context usefulness depends on transferability of changes

### Scaling Analysis
- 14B guidance outperforms 8B on saturated tasks (AHC058, TriMul)
- Larger guidance applies mechanisms with finer specialization: adapts fusion strategy to channel size, combines state-dependent valuation with beam-based initialization
- Probing shows 14B explains SOTA code mechanisms more accurately (3/3 correct vs. 1/3 for 8B on TriMul probes)

## Theoretical and Practical Implications

**Theoretical implications:**
- Establishes that test-time adaptation can be effectively directed at strategic guidance rather than full solution generation
- Demonstrates "learning in thought space" — natural-language proposals as learned actions — as a viable alternative to latent representations or complete programs
- Reveals that credit assignment improvement (separating strategy from implementation) yields measurable gains

**Practical implications:**
- **Cost reduction**: ~58% reduction in adaptation cost when using large frozen executors
- **Architecture insight**: Task decomposition (strategist + executor) mirrors human problem-solving and scales better than monolithic adaptation
- **Generalizability**: Works across heterogeneous domains (combinatorial optimization, heuristic programming, ML, GPU kernels) with identical framework

## Conclusion

Guidance-TTT demonstrates that separating strategic adaptation from solution implementation enables more efficient and effective test-time learning for scientific discovery. The framework produces state-of-the-art solutions across four distinct domains while substantially reducing computational cost.

**Future directions identified:**
1. Richer guidance representations beyond natural-language proposals
2. Improved credit assignment between guidance and execution models
3. More diversity-aware archive search strategies
4. Investigating how larger guidance models might further improve discovery on saturated tasks

---

_Markdown view of https://picx.dev/p/txYkn4, served by PicX — AI-generated visual whiteboard summaries of research papers._