PaperlayerView exact revision on arXiv

Source-linked AI summary

CompEvo: Competition-Induced Evolution for Multi-Agent in News-Driven Time Series Forecasting

Yuxuan Zhang, Yangyang Feng, Yong Guan, Daifeng Li, Kexin Zhang, Junlan Chen, Bowen Deng, Jun Liu, Zehua Zeng

arXiv:2609.09195v1cs.NEcs.MA

TL;DR

Multi-agent news-driven forecasting faces converging evidence-seeking strategies and limited theoretical grounding for strategy updates. CompEvo combines evolutionary-game analysis with trainable competitive strategy execution, selection, and evolution. Across four real-world benchmarks, it reports lower forecasting errors than a strong multi-agent baseline while maintaining diverse and specialized strategies.

  • Problem

    Multi-agent news-driven forecasting faces degeneration of thought, with agents converging on similar evidence-seeking behaviors, and insufficient theoretical grounding for strategy updates.

    Source: §1 · pp. 1–2

  • Method

    CompEvo formulates agent competition using evolutionary game theory and trains agents through strategy execution, fitness-based differentiable selection, and competition-induced strategy evolution.

    Source: §1 · p. 2

  • Results

    Across four real-world benchmarks, CompEvo reduces RMSE by 29.8% and MAPE by 30.3% relative to MAD-RL, while the paper reports improved forecasting accuracy and diverse, specialized strategies.

    Source: §5.2 · p. 7, §6 · p. 9

  • Takeaways & Limitations

    CompEvo maintains diverse and specialized agent strategies, with analyses linking its gains to fitness-guided specialization across backbones and evidence quality.

    Source: §5.6 · p. 9, §6 · p. 9

    Validation is on NTSF with directly relevant textual events; multimodal evidence and broader forecasting or decision-making settings remain future directions.

    Source: Limitations · p. 10

Abstract

from arXiv · show

News-driven time series forecasting uses evolving textual events together with historical observations to predict future values, supporting applications such as market risk monitoring and resource scheduling. In multi-agent settings, two challenges still remain. The first is degeneration of thought, where agents converge to similar evidence-seeking behaviors. The second is insufficient theoretical grounding, where strategy updates are often heuristic and lack a principled formulation. To address the above challenges, we propose CompEvo, a competition-induced evolution framework for multi-agent news-driven time series forecasting. For theoretical grounding, we introduce an evolutionary game formulation to guarantee equilibrium existence and optimization convergence. Building on this formulation, we construct a trainable multi-agent evolution framework that integrates strategy execution, fitness-based differentiable selection, and competition-induced strategy evolution. CompEvo enables heterogeneous agents to explore diverse news evidence, converts forecasting feedback into differentiable influence weights, and evolves agent strategies under competitive pressure to preserve effective logic while maintaining diversity. Experiments on four real-world datasets show that CompEvo reduces RMSE by 27.3% and MAPE by 26.2% on average over strong baselines. Further analysis indicates that CompEvo successfully maintains diverse and specialized agent behaviors.

1 Introduction

Multi-agent news-driven forecasting faces converging evidence-seeking strategies and limited theoretical grounding for strategy updates. CompEvo combines competitive strategy evolution with an evolutionary-game formulation.

1 Introduction

  • Multi-agent NTSF can suffer degeneration of thought, as agents imitate peers and converge on homogeneous evidence-seeking strategies.The introduction also identifies insufficient theoretical grounding when heuristic updates lack formal convergence and stability guarantees.Source: §1 · pp. 1–2
  • CompEvo uses strategy execution, fitness-based differentiable selection, and competition-induced evolution to preserve effective strategies while encouraging diverse evidence-seeking behaviors.Selection converts forecasting feedback into influence weights; strategy evolution uses partial peer information.Source: §1 · p. 2
  • The paper formulates agent interaction as an incomplete-information evolutionary game and analyzes equilibrium existence, local convergence, long-term regret, and fitness-weight monotonicity under stated assumptions.These analyses provide theoretical grounding for stable strategy retention under differentiable optimization.Source: §1 · p. 2
  • 29.8% lower RMSE and 30.3% lower MAPE than the strongest multi-agent baseline are reported across four real-world benchmarks.The implementation and data are also released to support reproducibility.Source: §1 · p. 2

2 Related Work

Prior work incorporates news into forecasting and uses agents to process evidence, while evolutionary approaches offer strategy adaptation. The paper identifies fixed or heuristic strategies and nondifferentiable operations as gaps.

2.1 LLMs for News-Driven Forecasting

  • Agentic forecasting methods retrieve, filter, refine, or discuss news, but often rely on fixed evidence-seeking strategies or consensus coordination.The paper notes that heterogeneous news-selection strategies can remain valid simultaneously and should evolve with forecasting feedback.Source: §2.1 · p. 2

2.2 Evolutionary Game-Theoretic Approaches

  • Existing LLM evolutionary frameworks use discrete or prompt-level operations, lacking differentiable selection that links forecasting feedback to strategy retention while preserving diverse evidence-seeking strategies.This gap motivates modeling news-driven forecasting as competition-induced evolution.Source: §2.2 · p. 2

3 Preliminaries

The preliminaries frame news-driven forecasting as a multi-agent competition over textual strategies, using historical observations and time-aligned public news to predict a future numerical value.

3 Preliminaries

  • NTSF combines historical numerical observations with a time-aligned public news pool to predict a future numerical value.The formulation denotes the historical window by x_{t−L:t−1} and the news pool by C_t.Source: §3 · p. 3
  • A population of heterogeneous agents applies distinct evidence-seeking strategies to shared news, producing customized news subsets and individual forecasts from identical historical data.Their differing strategies transform shared raw evidence into distinct forecasting behaviors.Source: §3 · p. 3

4 Methodology

CompEvo uses a shared LLM backbone with agent-specific adaptation, executes separate evidence-seeking strategies, selects agents using forecasting feedback, and evolves strategies using partial peer information.

4 Methodology

  • A shared LLM backbone with separate PEFT parameters supports agent-specific adaptation and heterogeneous textual logic.The paper distinguishes a parameterized policy-level strategy from its generated textual-logic realization.Source: §4 · p. 3

4.1 Strategy Execution

  • Strategy execution generates textual logic that filters external evidence and produces an individual forecast for each agent.Separate forecasts allow agent-level errors to inform selection and evolution rather than erasing specialization through early averaging.Source: §4.1 · p. 3

4.2 Fitness-based Selection

  • Fitness-based selection converts forecasting feedback into differentiable agent weights and uses long-term fitness to smooth selection across noisy rounds.Agents with higher fitness gain greater influence in aggregation and forecasting loss.Source: §4.2 · p. 4

4.3 Competition-Induced Strategy Evolution

  • Competition-induced strategy evolution updates textual logic using each agent’s local state and partial peer information, without fully replacing its own reasoning direction.The update fuses candidate logic with opponent-informed representations through gates.Source: §4.3 · p. 4

4.4 Training Objective

  • The training objective jointly optimizes forecasting, strategy updates, diversity, and pruning to discourage similar logic and weak-agent influence.Its terms evaluate aggregate forecasts, convert relative reward into logic updates, separate reasoning paths, and suppress persistently weak agents without eliminating them.Source: §4.4 · p. 5

4.5 Theoretical Justification

  • The theoretical analysis establishes a mixed-strategy Bayesian Nash equilibrium, supports local convergence, and shows influence weights rise with long-term fitness under soft selection.The regret analysis also finds the competitive update rule more favorable than debate-based interaction.Source: §4.5 · p. 5

5 Experiments

Across four real-world datasets, CompEvo improves on the multi-agent baseline and maintains strong performance across backbones, while ablations and analyses examine its optimization, diversity, and robustness. Results also show better behavior under news perturbations and clearer agent specialization.

5.1 Experimental Setup; 5.3 Backbone Robustness

  • Experiments cover Electricity, Exchange, Traffic, and Bitcoin; on Electricity, CompEvo remains effective across backbone models.Table 4 provides supplementary backbone-robustness results.Source: §5.1 · p. 5, Table 4 · p. 7, §5.3 · p. 7

5.2 Main Results

  • 29.8% lower RMSE and 30.3% lower MAPE than MAD-RL on the four-dataset macro average; CompEvo also beats SA and MAEL.The comparison reports relative error reductions over MAD-RL; the paper attributes gains over SA and MAEL to population-level competition rather than stronger individual agents or simple aggregation.Source: §5.2 · pp. 6–7, Table 3 · p. 7

5.4 Ablation Study

  • Every ablation variant underperforms; simplifying opponent-aware updates or using uniform weights produces the largest overall drops.A post-training check on 100 temporally later Electricity samples reports RMSE 333.38 and MAPE 6.99%; the authors say it argues against a memorization-only explanation but does not replace evaluation on continuously updated news streams.Source: §5.4 · p. 7

5.5 Optimization Stability and Efficiency

  • Figure 3 shows faster, steadier convergence than SA and MAD-RL; reward plateaus at a stable strategy, while three-agent costs do not estimate the 10-agent default.In the controlled three-agent setting, CompEvo uses 14.5 hours per epoch, 8.57 seconds per inference step, and 16.4 GB peak memory.Source: Figure 3 · p. 7, §5.5 · pp. 7–8, Table 7 · p. 8

5.6 Strategy-Update Diversity and Specialization

  • CompEvo maintains higher LUD and more unique news selections, with agents developing distinct topical focuses.Appendix checks report that LUD is associated with larger error reduction and that its ranking remains stable across models.Source: §5.6 · pp. 8–9, Figure 5 · p. 9

5.7 Hyperparameter Sensitivity; 5.8 Robustness under News Perturbations

  • Moderate regularization performs best (λdiversity = 0.1; λpruning = 0.01), and CompEvo degrades more slowly under shuffled, irrelevant, or noisy news.The regularization sweeps show representative results on Electricity and Exchange; similar perturbation patterns are reported on the other datasets.Source: §5.7 · p. 9, Figure 6 · p. 9, §5.8 · p. 9

5.9 Strategy Diversity Visualization

  • CompEvo’s final-epoch agent embeddings spread across distinct semantic areas, whereas MAD-RL’s cluster tightly.The authors connect the visualization to heterogeneous reasoning and specialization under diversity regularization and competitive evolution.Source: Figure 8 · p. 9, §5.9 · p. 9

6 Conclusions

CompEvo formulates multi-agent news-driven time series forecasting as competition-induced evolution, with a trainable objective inspired by evolutionary game theory. The paper reports improved accuracy alongside diverse, specialized strategies, with analyses linking gains to fitness-guided specialization.

6 Conclusions

  • CompEvo improves forecasting accuracy while maintaining diverse, specialized agent strategies, with analyses linking the gains to fitness-guided specialization robust across backbones and evidence quality.Source: §6 · p. 9
  • CompEvo frames agent competition as a unified differentiable objective, allowing agents to adapt through end-to-end training.Source: §6 · p. 9
  • Ablation, convergence, and perturbation analyses connect the reported gains to fitness-guided specialization.Source: §6 · p. 9

Limitations

The paper identifies boundaries in its validation scope and summarizes theoretical support for CompEvo, including equilibrium existence under stated conditions. It also defines notation for the forecasting task and agent strategies.

Limitations

  • CompEvo is validated on NTSF benchmarks where external textual events are directly relevant to future numerical values; broader forecasting and decision-making scenarios remain future directions.Source: Limitations · p. 10
  • The framework focuses on textual news, while incorporating image, video, audio, and structured event records is proposed as a way to broaden applicability.Source: Limitations · p. 10

A Notation

  • The notation describes forecasting from historical observations and time-aligned public news, with individual predictions produced by agents.Source: A Notation · p. 12

B Theoretical Justification

B.1 Existence of a BNE

  • A Bayesian Nash equilibrium exists under the stated conditions, including a localized quasi-concavity approximation of payoff with respect to action.Source: §B.1 · p. 13

B.2 Local convergence analysis

The analysis argues that minimizing CompEvo’s total objective converges to a stationary point interpretable as a local equilibrium, subject to regularity assumptions. It uses expected loss decrease and stochastic approximation to support this convergence account.

B.2 Local convergence analysis

  • Under compactness, smoothness, and bounded-variance assumptions, minimizing Ltotal converges to a stationary point corresponding to a local equilibrium of the induced game.Source: §B.2 · p. 13
  • The proof targets local rather than global convergence, citing the intractability of finding a global optimum for non-convex LLM value approximation in finite time.Source: §B.2 · pp. 13–14
  • The objective combines prediction loss, policy-gradient updates, and diversity and pruning strategy terms, with the regularizers encouraging diverse logic and suppressing incorrect logic interference.Source: §B.2 · p. 14
  • When strategy and main-objective gradients align, diversity may help escape local minima or saddle points, while pruning induces sparsity.Source: §B.2 · p. 14
  • The analysis states that diversity and pruning weights regulate strategy-gradient magnitude to mitigate gradient conflicts and dampen oscillations.Source: §B.2 · p. 14
  • Expected total loss decreases under the Lyapunov argument, and stochastic approximation yields stable convergence to a local minimizer when the stated assumptions hold.Source: §B.2 · p. 15
  • At a local minimum, no agent can improve its payoff through a local unilateral strategy update, supporting the local-equilibrium interpretation.Source: §B.2 · p. 15

B.3 Proof of Bayesian Regret

The analysis supports stable strategy improvement under the stated assumptions by establishing a sublinear Bayesian-regret bound for CompEvo.

B.3 Proof of Bayesian Regret

  • The value and Q functions represent expected rewards and future expected rewards, respectively, in the regret analysis.The Q function is defined over an agent’s state and actions, including the actions of other agents.Source: §B.3 · pp. 15–16
  • A performance-difference lemma is proved by recursively expanding value terms and using stationarity of the state distribution.The derivation then supports the Bayesian-regret bounding analysis.Source: §B.3 · pp. 16–17
  • CompEvo’s Bayesian regret is bounded sublinearly as T approaches infinity, supporting stable strategy improvement under the stated assumptions.The authors contrast this with linear regret in adversarial frameworks.Source: §B.3 · pp. 17–18
  • The regret proof decomposes Q-value differences into an estimation error and a policy-suboptimality term.The analysis treats the two terms separately before deriving the final bound.Source: §B.3 · pp. 17–18

B.4 Comparative Analysis: Linear Regret in Adversarial Frameworks

This section contrasts zero-sum debate with CompEvo’s evolutionary dynamics, then describes aggregation monotonicity, soft pruning, diversity regularization, and how the training losses divide their roles.

B.4 Comparative Analysis: Linear Regret in Adversarial Frameworks

  • A zero-sum debate framework has linear total regret because agents retain persistent single-step suboptimality.The paper contrasts this with CompEvo’s sublinear regret and argues that zero-sum dynamics do not support stable long-term strategy improvement.Source: §B.4 · p. 18

B.5 Aggregation Mechanism: Monotonicity of Fitness-based Weighting

  • When an agent’s long-term fitness increases, its aggregation weight is non-decreasing, assuming nonnegative gating and positive temperature.The paper presents this monotonicity as supporting the survival-of-the-fittest mechanism.Source: §B.5 · p. 19

C.1 Soft Pruning

  • Soft pruning uses a learnable gate and L1 penalty to smoothly suppress persistently weak agents while allowing later recovery.It does not eliminate agents discretely.Source: §C.1 · p. 19

C.2 Logic Diversity

  • The diversity penalty minimizes pairwise cosine similarity among candidate logics before opponent-aware mixing, encouraging distinct agent logics.Source: §C.2 · p. 20

C.3 Module Interaction Summary

  • The training loop assigns reward-linked logic updates to Lpg, diversity preservation to Ldiv, and differentiable selection to Lprune with gated softmax.Source: §C.3 · p. 20

D.1 Details of Datasets

The appendix documents benchmark data and implementation choices, gives an input-output example, and specifies how the baseline comparison is configured.

D.1 Details of Datasets

D.2 Implementation Details

  • CompEvo uses 10 agents with a shared frozen instruction-tuned backbone and agent-specific lightweight adapters; logic and forecasting adapters use rank 16.The implementation also uses GRPO to align logic generation with forecasting rewards.Source: §D.2 · p. 20

D.3 Forecasting Input Template for Fine-tuning LLM

  • The fine-tuning template pairs historical load data with a next-day VIC load request, date, holiday status, weather, and aligned news with rationales.It shows a numerical forecast as output.Source: §D.3 · p. 20

D.4 Baseline Details

  • Baseline methods share data splits, a 48-step horizon, metrics, validation folds, and candidate news pools, but their trainable parameter counts are not capacity matched.CompEvo uses 26.7% fewer trainable parameters than MAD-RL and 2.20× as many as MAEL or MAE-GPT.Source: §D.4 · p. 21, Table 11 · p. 21

E.1 Discussion on Memorization Risk

The appendix examines later-period electricity performance, news perturbations, hyperparameter and prompt sensitivity, and how logic updates are measured. Its findings include a leakage-controlled future-data check and robustness patterns, alongside a rubric that distinguishes structural updates from surface rewrites.

E.1 Discussion on Memorization Risk

  • On 100 2026 Electricity samples spanning February 19–March 30, CompEvo trained in the main experiments was run without further parameter updates, achieving RMSE 333.38 and MAPE 6.99%.Source: §E.1 · p. 22

E.2 Robustness Analysis under News Perturbations

  • Under shuffled, irrelevant, noisy, and original news, CompEvo degrades slowest in RMSE; the authors attribute the pattern to maintained diversity and specialized reasoning.Similar trends are reported for MAE, MSE, and MAPE.Source: Figure 9 · p. 23, §E.2 · p. 23

E.3 Additional Hyperparameter Sensitivity Analysis / E.4 Impact of Different Number of Initial Agents

  • Diversity and pruning regularization work best at moderate strength, while larger initial populations generally yield lower later-epoch MAPE on Electricity, with N = 10 best finally.The main experiments use 10 agents.Source: Figure 10 · p. 23, §E.3 · p. 23, Figure 11 · p. 23

E.5 Impact of Strategy Weight (λPG)

  • A moderate policy-update weight performs best: too little makes the method mostly fitness-weighted, while too much destabilizes forecasting.Source: §E.5 · p. 23, Table 13 · p. 24

E.6 Impact of LLM Generation Temperature / E.7 Aggregation Temperature (τ) / E.8 Role of EMA Smoothing Factor (β)

  • Moderate generation stochasticity supports exploration without large performance variance; aggregation temperature has a U-shaped sensitivity, and EMA smoothing trades responsiveness against stability.The selected settings are Tgen = 0.7 and τ = 0.5; small β reacts quickly but is noise-sensitive, whereas large β yields slower-moving fitness estimates.Source: §E.6 · p. 23, Figure 12 · p. 24, Table 14 · p. 24

E.9 Agent-side Instruction Robustness

  • Paraphrasing only the agent-side logic-update instruction yields nearly overlapping MAPE curves when the judge prompt, model, data, and hyperparameters stay fixed.The authors interpret this as evidence against brittleness to that wording choice.Source: Figure 14 · p. 24, §E.9 · p. 24

F.1 Overview / F.2 Formal Definition of LUD / F.3 Rubric Indicators: Detailed Criteria

  • Logic Update Degree combines embedding-based update magnitude with rubric indicators, and counts an update only when at least one innovation-related indicator is present.The indicators target strategy shifts and evidence recombination; paraphrases, stylistic changes, generic additions, and repeated reasoning are excluded.Source: §F.1 · p. 24, §F.2 · p. 25, §F.3 · p. 25

F.4 Validation of LUD

The section validates LUD by comparing judge assessments with human consensus and other LLM judges, and by testing its association with forecast-error reduction. It also documents the judging settings and logic-update prompts.

F.4 Validation of LUD

  • The LLM judge's decisions are consistent with human consensus, suggesting LUD reflects innovation-related structural updates rather than superficial paraphrasing.Three annotators labeled randomly sampled logic-update pairs using the two rubric indicators.Source: §F.4 · p. 25
  • Across 500 pairs, GPT-5.2 and Claude 4.6 Opus substantially agree with the original judge on categorical decisions and continuous LUD scores.The rubric and prompt template were held fixed.Source: §F.4 · p. 25, Table 15 · p. 26
  • On Electricity, LUD correlates positively with RMSE error reduction (r = 0.49, p < 0.001) across 1,200 update steps.The analysis associates larger, rubric-validated logic changes with forecasting improvement.Source: Table 16 · p. 26, §F.4 · p. 26

F.5 LLM judge settings

  • The judge uses temperature = 0 and a fixed prompt template, is blind to method identity, and receives only the logic pair.Source: §F.5 · p. 26

F.6 LLM Judge Prompt Template

  • The judge prompt compares two consecutive logics from the same agent and returns two binary indicators in JSON.Source: §F.6 · p. 26

G Instruction Templates

Related papers

Loading 2609.09195v1…