SMCEvolve: LLM-Driven Program Evolution

• SMCEvolve is a principled framework for LLM-driven program evolution that recasts program search as sampling from a reward-tilted distribution using SMC techniques.
• It integrates adaptive parent resampling and a mixture of mutation kernels with Thompson sampling to coordinate exploration and exploitation efficiently.
• Empirical evaluations across math, algorithm efficiency, and symbolic regression benchmarks demonstrate superior performance with reduced LLM calls compared to state-of-the-art methods.

SMCEvolve is a principled framework for LLM-driven program evolution that recasts program search as sampling from a reward-tilted target distribution and approximates that distribution with a Sequential Monte Carlo (SMC) sampler. In this formulation, parent selection, mutation, and stopping are not separate heuristics but coordinated components of a single probabilistic mechanism. The framework introduces adaptive parent resampling, a mixture of mutation with acceptance, and automatic convergence control, and it provides a finite-sample complexity analysis that bounds the LLM-call budget required to reach a target approximation error. Empirically, it is evaluated on math, algorithm efficiency, symbolic regression, and end-to-end ML research benchmarks, where it surpasses state-of-the-art evolving systems while using fewer LLM calls under self-determined termination (Jiang et al., 14 May 2026).

1. Variational formulation and target distribution

SMCEvolve begins from a task $q$, a program space $X$, an LLM prior over programs $p_0(x\mid q)$, and an evaluator $R(x)$ with bounded oscillation on the support of $p_0$. The objective is to find high-reward programs while remaining faithful to the LLM prior’s notion of plausible code or symbolic structure. This is expressed through a KL-regularized variational objective, $\max_p \mathbb{E}_{x\sim p}[R(x)] - D_{KL}(p\|p_0)$, whose exact solution is a reward-tilted distribution (Jiang et al., 14 May 2026).

The target distribution is

$$\pi(x)=p^*(x\mid q)=\frac{1}{Z(q)}\,p_0(x\mid q)\,\exp\big(\beta R(x)\big), \qquad Z(q)=\sum_{x\in X} p_0(x\mid q)\,\exp\big(\beta R(x)\big).$$

Here $\beta\ge 0$ is an inverse temperature that controls the strength of reward tilting. When $\beta$ is small, the search remains close to the prior; when $\beta$ is larger, the distribution concentrates more strongly on high-reward programs. The paper presents this distribution as the unique optimizer of the KL-regularized objective.

To make sampling tractable, SMCEvolve introduces an annealing path of intermediate distributions,

$X$0

The resulting search process is not framed as greedy improvement over a single incumbent, but as sequential approximation to a family of bridge distributions that gradually transform the prior into the reward-tilted target.

2. Sequential Monte Carlo design

The SMC construction supplies the operational core of SMCEvolve. Given forward mutation kernels and backward kernels, the general particle weight is

$X$1

where $X$2. The paper chooses $X$3 as the time reversal of $X$4 under $X$5, which cancels the forward and backward terms and yields a simplified importance weight depending only on the parent’s reward:

$X$6

This simplification is central: parent selection is derived from the annealing path rather than chosen ad hoc (Jiang et al., 14 May 2026).

Resampling is governed by the Effective Sample Size,

$X$7

SMCEvolve selects the next annealing factor $X$8 adaptively by bisection so that $X$9 for a threshold $p_0(x\mid q)$0. With $p_0(x\mid q)$1, the normalized parent weights are proportional to $p_0(x\mid q)$2. As $p_0(x\mid q)$3, these weights approach uniform exploration; as $p_0(x\mid q)$4 grows, they concentrate on higher-reward parents. Systematic resampling is used for efficiency and stability.

The paper further emphasizes diversity preservation. Early ESS thresholds and small $p_0(x\mid q)$5 keep bridges short and parent coverage broad, while later stages shift toward exploitation. Island-based parallelism with periodic migration is used to maintain diversity across subpopulations. This suggests that SMCEvolve treats exploration–exploitation trade-offs as a property of the annealing schedule itself rather than as a separately tuned heuristic.

3. Mutation kernels, acceptance, and automatic convergence control

Mutation is implemented as a mixture of LLM-driven proposal kernels,

$p_0(x\mid q)$6

The instantiation in the paper uses four proposal modes formed by crossing edit granularity with information source: local diff edit versus full rewrite, and with inspiration versus no inspiration. The resulting kernels are diff_with_inspo, diff_no_inspo, rewrite_with_inspo, and rewrite_no_inspo. Thompson sampling adapts the mixture weights $p_0(x\mid q)$7 by favoring kernels that recently produced accepted, high-reward proposals (Jiang et al., 14 May 2026).

To enforce invariance to $p_0(x\mid q)$8, the ideal Metropolis–Hastings rule would be

$p_0(x\mid q)$9

Because black-box LLM APIs do not expose proposal densities, SMCEvolve uses a reward-only approximation,

$R(x)$0

This always accepts non-decreasing reward moves and accepts decreases with probability that decays exponentially in the reward gap. Invalid or non-compiling proposals receive minimal reward, are typically rejected by the acceptance rule, but remain in the historical pool as future inspiration.

Automatic convergence control (ACC) ties annealing, stopping, and budget together. At each stage, the method chooses the largest $R(x)$1 such that $R(x)$2. If $R(x)$3, then $R(x)$4 and the algorithm stops; otherwise it continues. Termination therefore occurs when the annealing path reaches the target distribution rather than after a pre-fixed iteration count. The total LLM-call budget is

$R(x)$5

with $R(x)$6 particles, $R(x)$7 SMC stages, and $R(x)$8 one-step proposal/acceptance operations per particle per stage.

4. Finite-sample guarantees and their assumptions

A distinctive feature of SMCEvolve is that it does not present bounded growth or convergence as an empirical convenience alone. The paper states a finite-sample complexity theorem for the terminal empirical estimate

$R(x)$9

where $p_0$0 is any bounded test function with $p_0$1. Under an ESS-constructed schedule, adjacent bridge regularity, and mutation kernels that are $p_0$2-invariant and uniformly ergodic with rate $p_0$3, the paper gives conditions under which

$p_0$4

and the LLM-call budget satisfies

$p_0$5

where $p_0$6 is the reward oscillation and $p_0$7 hides polylogarithmic factors (Jiang et al., 14 May 2026).

The proof sketch specializes a theorem of Marion et al. (2023) to the reward-tilted path used by SMCEvolve. Three conditions are highlighted. First, bridge regularity: small $p_0$8 keeps adjacent distributions close enough for ESS control. Second, mutation mixing: uniform ergodicity yields a mixing cost $p_0$9. Third, path length: bounded reward oscillation implies $\max_p \mathbb{E}_{x\sim p}[R(x)] - D_{KL}(p\|p_0)$0, where $\max_p \mathbb{E}_{x\sim p}[R(x)] - D_{KL}(p\|p_0)$1. The stated regime requirements include $\max_p \mathbb{E}_{x\sim p}[R(x)] - D_{KL}(p\|p_0)$2, $\max_p \mathbb{E}_{x\sim p}[R(x)] - D_{KL}(p\|p_0)$3, and $\max_p \mathbb{E}_{x\sim p}[R(x)] - D_{KL}(p\|p_0)$4.

The paper also makes clear that some ingredients are idealized. The practical acceptance rule is reward-only because full LLM densities are unavailable, and the invariance and uniform ergodicity assumptions are approximated through kernel design and acceptance corrections. Thus the guarantees are principled but conditional on assumptions that are stronger than what can be directly verified for black-box LLM proposal mechanisms.

5. Empirical behavior across benchmark families

The empirical evaluation spans four benchmark families: ten math tasks from AlphaEvolve-style combinatorial geometry, eight algorithm-efficiency tasks from AlgoTune, sixteen symbolic regression tasks from LLM-SRBench, and AutoResearch on TinyStories pretraining. Baselines are REEVO, OPENEVOLVE, and SHINKAEVOLVE. For most domains the LLM ensemble is gpt-5-mini + gemini-3-flash, while AutoResearch uses gpt-5.4 + gemini-3-pro. Baseline budgets are fixed at 200 LLM calls for Math and AutoResearch, 400 for Symbolic Regression, and 500 for AlgoTune; SMCEvolve instead terminates automatically and has a mean call count strictly below those fixed budgets across all domains (Jiang et al., 14 May 2026).

On the math suite, SMCEvolve wins most tasks. Reported best rewards include $\max_p \mathbb{E}_{x\sim p}[R(x)] - D_{KL}(p\|p_0)$5 on Circle Packing $\max_p \mathbb{E}_{x\sim p}[R(x)] - D_{KL}(p\|p_0)$6, compared with $\max_p \mathbb{E}_{x\sim p}[R(x)] - D_{KL}(p\|p_0)$7 for Shinka and $\max_p \mathbb{E}_{x\sim p}[R(x)] - D_{KL}(p\|p_0)$8 for Reevo; $\max_p \mathbb{E}_{x\sim p}[R(x)] - D_{KL}(p\|p_0)$9 on Hexagon Packing $$\pi(x)=p^*(x\mid q)=\frac{1}{Z(q)}\,p_0(x\mid q)\,\exp\big(\beta R(x)\big), \qquad Z(q)=\sum_{x\in X} p_0(x\mid q)\,\exp\big(\beta R(x)\big).$$0, compared with $$\pi(x)=p^*(x\mid q)=\frac{1}{Z(q)}\,p_0(x\mid q)\,\exp\big(\beta R(x)\big), \qquad Z(q)=\sum_{x\in X} p_0(x\mid q)\,\exp\big(\beta R(x)\big).$$1 for OpenEvolve; $$\pi(x)=p^*(x\mid q)=\frac{1}{Z(q)}\,p_0(x\mid q)\,\exp\big(\beta R(x)\big), \qquad Z(q)=\sum_{x\in X} p_0(x\mid q)\,\exp\big(\beta R(x)\big).$$2 on Heilbronn Convex $$\pi(x)=p^*(x\mid q)=\frac{1}{Z(q)}\,p_0(x\mid q)\,\exp\big(\beta R(x)\big), \qquad Z(q)=\sum_{x\in X} p_0(x\mid q)\,\exp\big(\beta R(x)\big).$$3, compared with $$\pi(x)=p^*(x\mid q)=\frac{1}{Z(q)}\,p_0(x\mid q)\,\exp\big(\beta R(x)\big), \qquad Z(q)=\sum_{x\in X} p_0(x\mid q)\,\exp\big(\beta R(x)\big).$$4 for OpenEvolve and $$\pi(x)=p^*(x\mid q)=\frac{1}{Z(q)}\,p_0(x\mid q)\,\exp\big(\beta R(x)\big), \qquad Z(q)=\sum_{x\in X} p_0(x\mid q)\,\exp\big(\beta R(x)\big).$$5 for Shinka; and $$\pi(x)=p^*(x\mid q)=\frac{1}{Z(q)}\,p_0(x\mid q)\,\exp\big(\beta R(x)\big), \qquad Z(q)=\sum_{x\in X} p_0(x\mid q)\,\exp\big(\beta R(x)\big).$$6 on Kissing Number $$\pi(x)=p^*(x\mid q)=\frac{1}{Z(q)}\,p_0(x\mid q)\,\exp\big(\beta R(x)\big), \qquad Z(q)=\sum_{x\in X} p_0(x\mid q)\,\exp\big(\beta R(x)\big).$$7, compared with $$\pi(x)=p^*(x\mid q)=\frac{1}{Z(q)}\,p_0(x\mid q)\,\exp\big(\beta R(x)\big), \qquad Z(q)=\sum_{x\in X} p_0(x\mid q)\,\exp\big(\beta R(x)\big).$$8 for Shinka and $$\pi(x)=p^*(x\mid q)=\frac{1}{Z(q)}\,p_0(x\mid q)\,\exp\big(\beta R(x)\big), \qquad Z(q)=\sum_{x\in X} p_0(x\mid q)\,\exp\big(\beta R(x)\big).$$9 for OpenEvolve. Several autocorrelation tasks and Erdős Minimum Overlap show SMCEvolve matching or slightly exceeding the best baseline.

On algorithm efficiency, the pattern is more mixed but remains favorable overall. The paper reports $\beta\ge 0$0 speedup on psd_cone_projection, compared with $\beta\ge 0$1 for Shinka; $\beta\ge 0$2 on fft_convolution, compared with $\beta\ge 0$3 for Shinka; and $\beta\ge 0$4 on lu_factorization, compared with $\beta\ge 0$5 for OpenEvolve. One exception is polynomial_real, where Shinka attains $\beta\ge 0$6 and SMCEvolve attains $\beta\ge 0$7.

On symbolic regression, SMCEvolve wins most tasks by the $\beta\ge 0$8 reward. It is best on all four bio_pop_growth tasks with scores $\beta\ge 0$9, $\beta$0, $\beta$1, and $\beta$2; it wins three of four chem_react tasks with $\beta$3, $\beta$4, $\beta$5, and $\beta$6; and it reports strong results on matsci and phys_osc, including $\beta$7 on MatSci2, $\beta$8 on PO1, and $\beta$9 on PO3. On AutoResearch, the paper states that SMCEvolve achieves the highest final reward and auto-stops via ESS, with convergence curves showing a smooth rise and earlier termination than fixed-budget baselines.

The ablations, conducted on Circle Packing $\beta$0, are used to isolate the framework’s components. Adaptive parent resampling attains $\beta$1, whereas uniform resampling yields $\beta$2 and greedy single-parent selection yields $\beta$3. Single-kernel variants and uniform mixtures also degrade performance: diff-no-inspo and rewrite-no-inspo both produce $\beta$4, rewrite-with-inspo yields $\beta$5, and uniform mixtures reach at most $\beta$6, while the full four-kernel mixture with Thompson sampling reaches $\beta$7. A fixed $\beta$8 comparison shows that balance matters: $\beta$9 gives $X$00, $X$01 gives $X$02, and $X$03 gives $X$04.

6. Implementation, limitations, and terminological context

The implementation details are concrete and relatively compact. Typical settings use $X$05 particles per island, $X$06, $X$07 in common cases, and an ESS threshold $X$08. Resampling is systematic, log-sum-exp stabilization is applied to $X$09, and island-based execution may use multiple independent SMC chains with periodic migration of top-$X$10 particles. The inspiration pool contains top-$X$11 programs by reward plus “maximally diverse” exemplars chosen by text embedding distance, and it includes rejected candidates as well. Temperatures are approximately $X$12, maximum tokens are approximately $X$13, and the AutoResearch setting uses a single RTX A5000 with 60 seconds per candidate and reward $X$14 (Jiang et al., 14 May 2026).

The stated limitations are also specific. The acceptance rule is only an approximation to full Metropolis–Hastings because LLM densities are unavailable. Uniform ergodicity and exact invariance are idealized kernel assumptions. ESS bisection empirically enforces short bridges, but formalizing data-dependent schedules remains open. Large $X$15, large $X$16, and frequent evaluations can be compute-intensive. The paper proposes extensions including multi-objective search, reward-free or constraint-first settings, richer proposal families, and tighter complexity bounds.

A terminological ambiguity exists in contemporaneous arXiv usage. One 2026 paper uses SMCEvolve as the title of a Sequential Monte Carlo framework for principled scientific discovery (Jiang et al., 14 May 2026). A different 2026 paper, “SkillMAS: Coupling Skill Evolution with MAS Restructuring for Skill Co-Evolution,” uses SMCEvolve to denote “skill co-evolution” achieved by coupling skill evolution and MAS restructuring under verified-trace evidence (Pan et al., 10 May 2026). By contrast, the EvoMAC paper explicitly states that the term “SMCEvolve” does not appear anywhere in that work (Hu et al., 2024). Accordingly, the term can denote either a specific SMC-based program-evolution framework or, in a separate multi-agent systems context, a skill co-evolution paradigm; disambiguation depends on the cited source.