New Study Challenges Diminishing Returns in Long-Horizon LLMs

A recent study introduces CoreThink, a novel approach that significantly improves the performance of understanding-how-random-seeds-influence-convergence-and-divergence-in-language-models/”>understanding-trojan-attacks-in-large-language-models-a-new-study-on-inverting-trojans-and-defensive-strategies/”>large-language-models-a-practical-skimmable-guide-to-llms/”>large-language-models-best-practices-for-live-llm-inference-in-production/”>large language models (LLMs) on long-horizon tasks. This method challenges the common assumption of diminishing returns in extended LLM execution.

Key findings

CoreThink utilizes symbolic reasoning with explicit planning checkpoints to mitigate the diminishing returns observed in long-horizon tasks. On 200-step benchmarks, CoreThink increased success rates from 62% to 79%—a remarkable 17-point gain, demonstrating the sustained benefits of structured planning.

Other notable findings include:

• Humans outperform Monte Carlo methods and GPT-4 on long-horizon planning tasks (source: E. Jin, 2024).
• Open-source prompted LLM agents can effectively execute numerous tool calls without additional training (Source: Li et al., 2025c).
• CoreThink facilitates the generation of complex, offline task plans using LLMs (Source: K. Hori, 2025).

CoreThink’s Architecture

CoreThink integrates a discrete symbolic planner that acts as the core for long-horizon reasoning. This planner constructs and maintains a task graph, incorporating explicit checkpoints at intervals of 50–100 steps. These checkpoints prevent the planning process from deviating too far from the intended trajectory, allowing for course correction.

The symbolic layer meticulously tracks intermediate goals, constraints, and tool actions. It validates progress at each checkpoint before proceeding to subsequent planning phases, ensuring plan fidelity and robust execution.

Benchmark Results

These results showcase how a discrete symbolic layer enhances both the reliability and trajectory adherence of CoreThink in long-horizon tasks.

Addressing Diminishing Returns

Long tasks often expose a critical weakness in LLMs: without an explicit plan, they may drift, repeat actions, or lose track of their objectives. CoreThink effectively counters this by implementing structured reasoning and establishing rollback points. This allows the model to stay on track and gracefully recover from errors.

Key Features

• Structured reasoning with rollback points: CoreThink employs a stepwise approach with clearly defined checkpoints, enabling rollbacks to prior states when deviations occur.
• Plan verification and checkpointing: Each step undergoes verification, preserving context and facilitating faster recovery from errors.
• Compute efficiency and horizon extension: CoreThink maintains modest planning overhead while significantly expanding the usable horizon, typically by 1.5 times.

Empirical Methodology

To rigorously evaluate CoreThink, a benchmark suite was developed, encompassing 100–200 step reasoning tasks across diverse domains. These tasks involve multi-step mathematics, tool-enabled planning, and uncertain inference, and were designed to assess both the correctness of results and the quality of intermediate steps.

Baselines included Monte Carlo-based search and GPT-4. Evaluation adhered to a human-in-the-loop methodology, similar to (Source: E. Jin, 2024), with human judges assessing both correctness and plan fidelity.

Results

These findings suggest a promising approach: combining human-like planning, robust error control, and prompt-driven LLMs for complex, long-horizon tasks.

Reproducibility

The CoreThink project prioritizes reproducibility, providing open-source code, data, and evaluation tools. Users can easily replicate the results by cloning the repository, setting up the Docker environment, and running the provided scripts.

Practical Implementation

Integrating CoreThink with an LLM involves defining the task graph, initializing CoreThink, executing with verification and rollback, optimizing tool calls, and referencing provided patterns in the repository. A step-by-step guide is provided in the repository.

Comparison and Evaluation

A comparison between CoreThink, baseline LLMs and Monte Carlo methods with replanning highlights CoreThink’s strengths and weaknesses.

Conclusion

CoreThink offers a significant advancement in addressing the challenges of diminishing returns in long-horizon LLM execution. Its structured approach, coupled with robust error control and reproducibility features, presents a compelling solution for tackling complex tasks.

Watch the Official Trailer