Do self-improving AI systems actually get smarter, or do they just win by being given more compute to try?
How does test-time search budget compare to evolution gains under matched conditions?
This explores whether self-improving systems (evolved agents, evolved harnesses, evolutionary search) really beat plain test-time search once both get the same compute and feedback, or whether they mostly benefit from spending more.
This explores whether self-improving systems (evolved agents, evolved harnesses, evolutionary search) really beat plain test-time search once both get the same compute and feedback, or whether they mostly benefit from spending more. The corpus doesn't hold one clean head-to-head number for this. It does hold a strong argument about why the comparison matters, and a few partial answers.
The first point is that compute alone buys a lot. Agentic deep research shows search budget scaling the same way reasoning tokens do, with steady gains that eventually flatten Does search budget scale like reasoning tokens for answer quality?. At the multi-agent level, about 80% of performance variance comes from token budget rather than clever coordination How does test-time scaling work at the agent level?. So when an evolved system beats a baseline, extra spending is the default explanation. The corpus makes this explicit for harness evolution: an evolved harness has to be compared against task-level test-time search with equal feedback and equal inference budget, and only the gap beyond that is credit to the design How should we measure gains from automatic harness evolution?.
The clearest matched result is one where evolution is itself the search method. Mind Evolution uses LLM-written crossover and mutation across separate populations, and it solves over 98% of planning tasks. It clearly outperforms best-of-N and sequential revision, which are the standard ways to spend inference compute Can evolutionary search beat sampling and revision at inference time?. Here the win comes from how the budget is used (recombining partial solutions), not from having more of it.
The other evidence is about evolving the agent itself, and it is weaker on this question. The Darwin Gödel Machine reaches 2.5× on SWE-bench by keeping an archive of agent variants and testing them empirically Can AI systems improve themselves through trial and error?. An evolved AIDE agent matches its human-built counterpart on four held-out benchmarks Does automated evolution match human-built agent performance?. Both are impressive, but their summaries don't say whether a matched-budget search baseline was tried. That is the missing control the harness-evolution note asks for.
Whether evolution pays off also depends on the model feeding it. Search needs varied, competent candidates to recombine. Training for diversity lets evolutionary search solve problems that entropy-collapsed policies can't reach at all Should training maximize diversity when models feed into search?. RL training on search agents narrows exploration in the same way, while SFT on diverse demonstrations keeps it wide Does reinforcement learning squeeze exploration diversity in search agents?. A matched-budget comparison therefore also measures how much diversity the base model has. Extra compute can't fully make up for a trained-in limit either How should we spend compute at inference time?, which fits the view that internal capability-building and external search complement each other rather than compete How do internal and external test-time scaling compare?.
Sources 10 notes
Mind Evolution, an evolutionary search strategy using LLM-generated crossover and mutation with island model diversity, solves 98%+ of planning tasks and significantly outperforms best-of-N and sequential revision strategies while working directly in natural language without task formalization.
DGM replaces formal proofs with empirical benchmarking and maintains an evolutionary archive of agent variants, achieving 2.5× improvement on SWE-bench and 2.2× on Polyglot by discovering capabilities like better code editing and context management.
Agentic deep research shows monotonic-to-diminishing-returns curves for search iterations, matching reasoning token scaling. This creates a new inference-compute axis: models can trade off reasoning budget against search budget to optimize answer quality.
Research shows test-time scaling methods split into internal (training models for autonomous reasoning) and external (inference-time search and verification). They complement rather than compete; internal builds capability while external extracts performance from existing capability.
Vector Policy Optimization trains models to emit varied competent solutions rather than converging to one answer. This unlocks search procedures like evolutionary algorithms to explore and combine modes, solving problems that entropy-collapsed policies cannot reach at all.
Show all 10 sources
Automatic harness evolution must be compared against task-level test-time search under equal feedback and inference budgets. Only gains beyond what matched search achieves are attributable to the harness design itself, not just more computation.
RL training compresses behavioral diversity in search agents through the same entropy collapse mechanism documented in reasoning—policies converge on narrow reward-maximizing strategies. SFT on diverse demonstrations preserves exploration breadth, suggesting diversity-preservation techniques are essential for RL search scaling.
Research shows that uniform inference budgets waste compute; allocation should vary by prompt. Test-time compute can substitute for training-time scaling on hard problems, but cannot overcome fundamental limitations set by the training regime.
AIDE85, evolved through seven accepted rewrites in 8 days, equals or surpasses AIDEhuman on four held-out benchmarks spanning in- and out-of-distribution tasks including weather forecasting. The result shows automated design iteration can match human-driven R&D on generalization.
Research shows 80% of multi-agent performance variance comes from token budget, not coordination intelligence. LatentMAS and shared-KV-cache approaches offer ways to decouple performance gains from token costs.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- Does Thinking More always Help? Understanding Test-Time Scaling in Reasoning Models
- DarwinX: Evolving Agent Harnesses Through Natural Selection
- Scaling LLM Test-Time Compute Optimally can be More Effective than Scaling Model Parameters
- Retrieval-augmented reasoning with lean language models
- When More Thinking Hurts: Overthinking in LLM Test-Time Compute Scaling
- Vector Policy Optimization: Training for Diversity Improves Test-Time Search
- Single-Agent LLMs Outperform Multi-Agent Systems on Multi-Hop Reasoning Under Equal Thinking Token Budgets
- Rethinking the Evaluation of Harness Evolution for Agents