EvoCast: Reliable Autonomous Research Agents for Iterative Forecasting Architecture Evolution
TLDR
EvoCast introduces autonomous research agents with cognition-authority separation to iteratively evolve time-series forecasting architectures, outperforming baselines in three real-world cases.
Reasoning
The paper presents a novel system combining LLM-driven hypothesis generation with deterministic program authorities, enabling reliable autonomous architecture evolution. Its strengths include evidence-grounded iterative refinement and real-world evaluation, but the abstract provides limited details on baseline comparisons and generalizability beyond forecasting.
Read-first score
Read-first score 76.8, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 64.
Field roles
Rank sensitivity
Stability: volatile; rank range: 40.
Keyword Scores
Deep Analysis
Innovations
- EvoCast: a fully autonomous research-agent system for iterative forecasting architecture evolution.
- Cognition-authority separation: LLM agents handle open-ended hypothesis generation and code implementation, while deterministic program authorities control source-edit boundaries, canonical evaluation, and promotion decisions.
- Evidence-grounded research direction generation using dataset characteristics, diagnostic results, prior rounds, and failure records.
- Task-specific baseline establishment and diagnosis through executed mechanism ablations, with experimental outcomes accumulated as evidence to guide subsequent rounds.
Methodology
EvoCast first establishes and diagnoses a task-specific baseline through executed mechanism ablations, then iteratively generates evidence-grounded research directions from dataset characteristics, diagnostic results, prior rounds, and failure records. Its design separates cognition (LLM agents for hypothesis generation and code implementation) from authority (deterministic program authorities for source-edit boundaries, canonical evaluation, and promotion decisions). It is evaluated on three real-world forecasting cases against selected baselines, strong forecasting models, and agent baselines, measuring implementation success, agent-side token/time cost, and forecasting performance.
Key Results
EvoCast completes complex architecture modifications with higher implementation success and lower agent-side token/time cost, and develops task-specific architectures that outperform selected baselines, strong forecasting models, and agent baselines in three real-world forecasting cases.