Researchers present ABBEL framework that replaces recursive summaries with belief states

The preprint released at the end of July introduces ABBEL, a framework that substitutes recursive summarization with belief-state updates in order to address the compression failure that large language models exhibit on long-term tasks. The authors demonstrate that models performing Context Compaction during reinforcement training do not close the performance gap to models that retain the full context, even on the relatively simple benchmark Combination Lock, a Wordle-style game that allows up to 16 guesses. Model-serving systems such as Cursor continue to advise users to avoid compression in the middle of coding sessions, and the researchers attribute the difficulty to the lack of high-quality human simulators for training (Lin et al., 2025; Tomlin et al., 2025).

The need for compression stems from a basic limitation: context windows cannot grow indefinitely when tasks such as software development require hundreds or thousands of interaction steps. The heuristic approach to date has been recursive summarization; Cursor Composer 2.5 employs compression during training (Cassano et al., 2026), and Grandcode of DeepReinforce—the first system to consistently defeat human competitors in online coding contests while using Qwen 3.5-397B—was also forced to embed context summarization. In practice, summarization models improve during reinforcement training but still fall short of full-context models, as illustrated by the average-attempts curve for Combination Lock.

ABBEL (Acting Through Belief Bottlenecks) changes the paradigm: instead of asking the model to produce a free-form textual summary of the interaction history, it formulates the summary as a belief state that the model updates cyclically on the basis of new information, following a recursive Bayesian estimation scheme. The belief state replaces the complete interaction history as the agent’s working context, and the update is performed by encoding the previous belief, the action taken, and the new observation into the next belief.

The central novelty is Belief Grading, an auxiliary reinforcement task that rewards the model for its ability to reconstruct selected information from the history out of the compressed belief state. The authors illustrate the heuristic in the coding domain: a good belief is shorter yet more faithful to the ability to reconstruct a git diff; balancing brevity and fidelity yields an efficient belief. In domains where task-specific heuristics are difficult to define, they propose a general autoencoder-inspired grading function that … (the source text is truncated here).

The paper points to a promising direction for the long-standing compression problem in long-term coding agents, but it remains a preprint without peer review. Full performance metrics against established baselines such as SWE-bench or multi-step coding benchmarks have not been released, and it is unclear how the mechanism copes with hallucinations within the belief state itself. Should the autoencoder-based grading prove effective in practice, future generations of tools like Cursor or Grandcode may incorporate it, yet until comprehensive results and open-source code are published, the contribution remains theoretical.