Tree-Structured Thought Evolution (TreEvo)

• Tree-Structured Thought Evolution (TreEvo) is an LLM-guided framework that evolves candidates as hierarchical thought trees for compositional alpha mining.
• It applies semantic-level operators like crossover, mutation, and pruning to modify coherent reasoning components rather than raw text.
• Empirical evaluations on datasets such as CSI300 and CSI500 show up to a 20% improvement in Information Coefficient over leading baselines.

Tree-Structured Thought Evolution (TreEvo) is an LLM-driven evolutionary framework in which candidate ideas are represented and evolved solely at the thought level, but those thoughts are represented as hierarchical trees rather than linear text (Ren et al., 22 Aug 2025). In the TreEvo formulation, the unit of evolution is not a chain of sentences, a code string, or a conventional Tree-of-Thought search branch. Instead, a candidate is a tree-structured thought whose internal subtrees correspond to coherent reasoning components, and whose root encodes the overall alpha expression to be mined in quantitative finance. The framework was introduced for alpha mining, where candidate signals are inherently compositional and tree-shaped, and it advances the claim that the structure of the evolved thought representation should match the structure of the target solution space (Ren et al., 22 Aug 2025).

1. Definition and conceptual position

TreEvo was introduced in the context of alpha mining, defined as the automatic discovery of signals (“alphas”) that can predict future asset returns and therefore support quantitative trading (Ren et al., 22 Aug 2025). The paper explicitly frames alpha mining as a kind of tree-based symbolic regression: a candidate alpha is a compositional formula built from market features such as open, high, low, close, volume, and VWAP, combined with operators into an interpretable formula (Ren et al., 22 Aug 2025). Prior symbolic approaches are described as suffering from computational inefficiency and dependence on prior knowledge, while recent LLM-driven evolutionary algorithms are described as promising but limited when they represent thoughts as plain-text prompts rather than hierarchical structures (Ren et al., 22 Aug 2025).

TreEvo’s central distinction is that it does not treat a thought as a linear reasoning trace. The paper explicitly contrasts TreEvo with Chain-of-Thought and Tree-of-Thought reasoning, stating that, in contrast to those methods, “TreEvo regards a whole tree as a thought” (Ren et al., 22 Aug 2025). This makes TreEvo neither a standard tree search procedure nor a code-only evolutionary algorithm. It is a thought-only evolutionary framework whose representation is hierarchical from the outset.

The immediate target problem is alpha mining, but the paper also presents a broader design principle: if the target solution space is inherently hierarchical, the thought representation used by LLM-guided search should also be hierarchical (Ren et al., 22 Aug 2025). This suggests that TreEvo is not merely a finance-specific heuristic; it is a representational proposal about how LLM-guided evolution should be aligned with compositional search spaces.

2. Tree-structured thought representation

TreEvo formalizes a thought as a tree

$\mathcal{T} = (V, E)$

where $V$ is the set of reasoning units, $E \subseteq V \times V$ defines hierarchical dependencies, each node $v_i \in V$ is a semantic or computational subcomponent, and the root node $v_0$ encodes the overall alpha expression (Ren et al., 22 Aug 2025).

This representation creates a strict separation between thought space and execution space. In thought space, the evolutionary algorithm maintains and modifies tree-structured thoughts. In execution space, the LLM translates each thought tree into executable code, the code is run on market data, and the resulting predictive performance becomes the candidate’s fitness (Ren et al., 22 Aug 2025). The framework therefore does not directly evolve code as its primary representation.

The paper’s illustrative thought tree is:

• Volume Weighted Close-to-Open Return
  • Calculate Close-to-Open Return
  • Use Close price
  • Use Open price
  • Weight by Volume
  • Use Volume
  • Multiplied by Return

This example is used to show how the root encodes the overall alpha idea, intermediate nodes encode semantic subideas, and leaves encode basic inputs or composition components (Ren et al., 22 Aug 2025). The claimed benefit is representational alignment: alpha expressions are naturally compositional, and TreEvo mirrors that with a thought tree in which leaves correspond to basic units, intermediate nodes correspond to meaningful subcomputations, and the root corresponds to the overall alpha concept (Ren et al., 22 Aug 2025).

A direct consequence of this design is that evolutionary edits can be made at the level of semantically meaningful subtrees rather than arbitrary token spans. The paper contrasts this with flat thought prompts, where reasoning components are entangled and crossover or mutation on text is less structure-aware (Ren et al., 22 Aug 2025).

3. Evolutionary process and operators

TreEvo uses a three-step framework: initialization, iterative evolution, and output selection (Ren et al., 22 Aug 2025). Initialization begins with a population $P$ of $N$ thoughts. The LLM is prompted with Initialization Prompts to generate $N$ initial tree-structured thoughts, then generates $N$ executable codes, one from each thought, and each code is executed and evaluated on market data (Ren et al., 22 Aug 2025).

The evolutionary process repeats until a stop criterion is met, such as a time or evaluation budget. In each round, TreEvo generates $N$ new thoughts by crossover, $V$0 by mutation, and $V$1 by pruning, producing $V$2 offspring thoughts (Ren et al., 22 Aug 2025). Each offspring is translated into executable code, run on the real-market dataset, and scored by predictive performance. Selection then chooses the best $V$3 individuals to update the population (Ren et al., 22 Aug 2025). The system returns the code corresponding to the best-performing thought found during the whole run.

The paper defines three semantic-level evolutionary operators:

Crossover

Given two thought trees $V$4 and $V$5, crossover is defined as subtree exchange:

$V$6

where $V$7, $V$8, and $V$9 denotes substituting the subtree at $E \subseteq V \times V$0 with the subtree at $E \subseteq V \times V$1, or more generally exchanging these subtrees (Ren et al., 22 Aug 2025). The paper’s prompt sketch states that crossover should “generate a new thought tree that synthesizes reasoning components from both parents while maintaining overall coherence” (Ren et al., 22 Aug 2025).

Mutation

Mutation selects a node $E \subseteq V \times V$2 in a thought tree $E \subseteq V \times V$3 and replaces its subtree with a newly generated tree $E \subseteq V \times V$4:

$E \subseteq V \times V$5

The notation is reported as slightly malformed in the paper, but the intended meaning is subtree replacement at node $E \subseteq V \times V$6 (Ren et al., 22 Aug 2025). Mutation is therefore a localized semantic edit: most of the idea is preserved, but one reasoning component is changed.

Pruning

Pruning identifies a redundant or semantically neutral subtree rooted at $E \subseteq V \times V$7, and replaces it with a simpler expression $E \subseteq V \times V$8:

$E \subseteq V \times V$9

Again, the notation is reported as slightly malformed in the paper, but the intended meaning is subtree simplification (Ren et al., 22 Aug 2025). The stated goal is to produce “a more concise thought that retains the core reasoning idea.”

The paper does not define a larger operator catalog with separate hierarchy-aware insertion, depth constraints, or explicit semantic consistency formulas (Ren et al., 22 Aug 2025). The core claim is narrower: crossover, mutation, and pruning become semantically meaningful when applied to thought subtrees rather than raw syntax.

4. Objective functions and empirical results

TreEvo evaluates candidate alphas primarily with Information Coefficient (IC) and Rank IC (Ren et al., 22 Aug 2025). IC is defined as the average daily Pearson correlation between predicted scores $v_i \in V$0 and future returns $v_i \in V$1:

$v_i \in V$2

where $v_i \in V$3 are predicted alpha values, $v_i \in V$4 are ground-truth future 5-day returns, and $v_i \in V$5, $v_i \in V$6 are daily cross-sectional means (Ren et al., 22 Aug 2025). The paper states that IC is also adopted as the training loss or search objective for the compared algorithms. Rank IC is defined as

$v_i \in V$7

and is used as an auxiliary evaluation metric (Ren et al., 22 Aug 2025).

The experiments use four real-market datasets across two markets: CSI300, CSI500, SPX, and DJI (Ren et al., 22 Aug 2025). The six raw features are opening price, closing price, highest price, lowest price, trading volume, and VWAP; all price and volume data are forward-adjusted for corporate actions (Ren et al., 22 Aug 2025). The chronological split is Training: 2016/01/01 – 2020/01/01, Validation: 2020/01/01 – 2021/01/01, and Test: 2021/01/01 – 2024/01/01 (Ren et al., 22 Aug 2025). For all LLM-driven evolutionary algorithms, the LLM is Qwen3-Max and the population size is 10 (Ren et al., 22 Aug 2025).

The search budget is deliberately small: TreEvo, EoH, and ReEvo are given 200 evaluations; XGBoost, LightGBM, AlphaGen, and QFR are given at least 1000 evaluations; and GP is given 20000 evaluations (Ren et al., 22 Aug 2025). Performance is averaged over 5 independent runs (Ren et al., 22 Aug 2025).

On CSI300 and CSI500, TreEvo achieves the best IC among the listed baselines. On CSI300, the reported IC values are XGBoost 0.0192, LightGBM 0.0158, GP 0.0445, AlphaGen 0.0500, QFR 0.0588, and TreEvo 0.0615 (Ren et al., 22 Aug 2025). On CSI500, the reported IC values are XGBoost 0.0173, LightGBM 0.0112, GP 0.0557, AlphaGen 0.0544, QFR 0.0708, and TreEvo 0.0742 (Ren et al., 22 Aug 2025). The paper states that even relative to the strongest baseline on each index, TreEvo improves IC by at least 4.59% on CSI300 and 4.80% on CSI500 (Ren et al., 22 Aug 2025).

Against LLM-driven evolutionary algorithms, the reported comparisons are also favorable on most datasets. On CSI300, EoH has IC 0.0506 and RankIC 0.0462, ReEvo has IC 0.0411 and RankIC 0.0409, and TreEvo has IC 0.0615 and RankIC 0.0649 (Ren et al., 22 Aug 2025). On CSI500, EoH has IC 0.0584 and RankIC 0.0653, ReEvo has IC 0.0426 and RankIC 0.0489, and TreEvo has IC 0.0742 and RankIC 0.0793 (Ren et al., 22 Aug 2025). On SPX, EoH has IC 0.0504 and RankIC 0.0458, ReEvo has IC 0.0362 and RankIC 0.0308, and TreEvo has IC 0.0403 and RankIC 0.0543 (Ren et al., 22 Aug 2025). On DJI, EoH has IC 0.0514 and RankIC 0.0546, ReEvo has IC 0.0368 and RankIC 0.0393, and TreEvo has IC 0.0601 and RankIC 0.0573 (Ren et al., 22 Aug 2025). The paper summarizes this pattern by stating that TreEvo improves IC by around 20% on CSI300, CSI500, and DJI, while on SPX it is behind EoH on IC but ahead on RankIC by 18.6% (Ren et al., 22 Aug 2025).

The paper also reports a trading case study on CSI300 under a Top-50/Drop-5 daily rebalancing strategy, where TreEvo achieves 16.91% improvement in cumulative return over the strongest baseline and more than 100% excess return relative to the CSI300 benchmark (Ren et al., 22 Aug 2025). In the abstract and introduction, the efficiency claim is summarized as 5x–100x acceleration, corresponding to the fact that TreEvo uses a much smaller evaluation budget than several traditional baselines (Ren et al., 22 Aug 2025).

5. Ablations, relation to Tree-of-Thought, and adjacent methods

Ablation results are central to the paper’s argument. The authors define TReEvo, a variant of ReEvo that uses the proposed tree-structured thoughts but keeps ReEvo-style code evolution operators (Ren et al., 22 Aug 2025). On averages over three runs, ReEvo IC is 0.0405 on CSI300 and 0.0468 on CSI500, while TReEvo IC is 0.0546 on CSI300 and 0.0631 on CSI500, corresponding to improvements of $v_i \in V$8 on CSI300 and $v_i \in V$9 on CSI500 (Ren et al., 22 Aug 2025). TreEvo then improves over TReEvo, with IC 0.0601 on CSI300 and 0.0726 on CSI500, corresponding to improvements of $v_0$0 on CSI300 and $v_0$1 on CSI500 (Ren et al., 22 Aug 2025). The paper’s conclusion is that the tree-structured thought representation is the dominant contributor, while compatible semantic-level operators further strengthen performance.

TreEvo is closely related to Tree-of-Thought methods, but it is not a Tree-of-Thought search procedure in the usual sense. Novelty-pruned Tree-of-Thought search adds a global memory-dependent diversity test to tree search and can be characterized as tree search with a diversity gate rather than a full population-based evolutionary loop (Hamm et al., 7 May 2026). Multi-agent ToT with a Thought Validator adds validator-based pruning and validated majority voting, but it does not implement mutation, crossover, adaptive population updates, or persistent branch inheritance (Haji et al., 2024). iToT introduces a human-visible and human-steerable reasoning workspace in which users can select branches, inject custom thoughts, and use semantic grouping, but it does not provide a formal evolutionary algorithm (Boyle et al., 2024). These systems share branch generation, filtering, or correction mechanisms, yet TreEvo differs in a more fundamental way: it makes the whole hierarchical thought tree the evolvable individual.

A later adjacent development is TreeCUA, which organizes GUI trajectories into an exploration tree with local verification, replay-based node reuse, adaptive depth-breadth balancing, and branch-derived preference learning (Jiang et al., 10 Feb 2026). TreeCUA is framed in GUI automation rather than alpha mining, and its nodes are executable GUI states rather than abstract thought states. Even so, its emphasis on shared-prefix tree growth, local verification, pruning, backtracking, and sibling-branch preference construction indicates that tree-structured evolutionary principles can also be deployed outside symbolic finance tasks (Jiang et al., 10 Feb 2026). This suggests that TreEvo belongs to a broader family of methods in which tree structure is used not only for search, but also for reuse, selection, and supervision.

6. Limitations, open questions, and significance

TreEvo’s main acknowledged limitation is that search has not fully converged within the experimental budgets (Ren et al., 22 Aug 2025). The paper notes that upward convergence curves suggest additional performance may still be available, but also emphasizes a practical trade-off: stopping too early may yield underdeveloped factors, while running too long may overfit (Ren et al., 22 Aug 2025). The authors suggest that future work should incorporate regularizations, complexity controls, or early stopping heuristics (Ren et al., 22 Aug 2025).

The paper does not deeply analyze how sensitive performance is to prompt design, how stable thought-to-code translation is, how often LLM-generated code fails or requires sanitization, or how tree validity is enforced in practice beyond prompting (Ren et al., 22 Aug 2025). A plausible implication is that TreEvo’s practical robustness depends not only on the representational idea of tree-structured thoughts, but also on the reliability of the LLM as both a semantic evolution agent and a code generator.

TreEvo’s broader significance lies in its proposal to change the object being evolved. Earlier LLM-guided evolutionary approaches evolved code only, or evolved linear thoughts plus code. TreEvo instead evolves hierarchical reasoning ideas solely at the thought level, and then renders those thoughts into executable artifacts for evaluation (Ren et al., 22 Aug 2025). The paper’s empirical argument is that this representation-and-operator alignment leads to better alpha quality, faster convergence, lower compute budget, and less dependence on manual expert engineering (Ren et al., 22 Aug 2025).

The method therefore occupies a specific position in the landscape of tree-based LLM reasoning. It is more explicitly evolutionary than Tree-of-Thought search, more structurally aligned than flat-prompt thought evolution, and more semantically grounded than symbol-level genetic programming (Ren et al., 22 Aug 2025). Its defining claim is not that all tree-based reasoning should become TreEvo, but that when the target hypothesis space is itself hierarchical, evolving tree-structured thoughts can be a better match than evolving linear prompts or raw code.