ABBEL framework replaces context summarization with belief states for long-horizon AI
New belief-grading method outperforms recursive summarization in long coding tasks without context bloat.
For LLMs to assist in complex tasks like software development spanning hundreds or thousands of steps, maintaining full interaction history in context is impractical. Current heuristic approaches rely on recursive summarization (context compaction), used by systems like Cursor's Composer 2.5 (Cassano et al., 2026) and Grandcode (DeepReinforce et al., 2026, first to beat all humans in online coding competitions). Yet even with models like Qwen 3.5-397B, compaction introduces a performance gap that never closes during RL fine-tuning, as demonstrated in the Combination Lock game. Model servers still advise users to avoid compaction mid-task.
ABBEL addresses this by treating summaries as belief states updated periodically, isolating belief generation from action selection. It introduces belief grading—an auxiliary RL task with heuristics (e.g., shorter is better, but git-diff reconstructability scores higher) or a general autoencoding-inspired objective that rewards the model for reconstructing recent observations from the belief state. This supervised bottleneck improves summarization quality without requiring massive high-quality training data, making long-horizon AI assistants more reliable.
- Recursive summarization (used in Composer 2.5, Grandcode) creates a persistent performance gap vs. full context, even after RL training.
- ABBEL uses belief states as compressed representations, updated via periodic prompts and isolated from action selection.
- Belief grading with reconstruction-based rewards improves summarization quality without needing large, curated datasets.
Why It Matters
ABBEL could unlock reliable long-horizon AI assistants for coding and multi-step reasoning without context limits or performance loss.