Improving the performance of RLM for Long-Context Reasoning with the help of structured graph state and combining Mamba SSM

Reasoning performance in language models degrades with long input contexts, even in frontier models. Recursive language models (RLMs) mitigate this by processing inputs in chunks, but they face a control problem; deciding when to stop, consolidate, or continue exploring. We introduce the RST (Recursive -Structured-State -Termination) Reasoning Engine, an adaptive-recursion architecture that addresses this challenge. RST represents reasoning as operations on a structured graph state -expansion (integrating new information) and consolidation (refining existing information). Their non-commutativity is quantified by an order-gap, Ω, which we track through a graph-spectral proxy. The proxy is derived as a first-order approximation to the canonical order-gap; rather than proving the two equivalent, we test empirically whether it tracks convergence well enough to guide adaptive recursion. We evaluate RST on three benchmarks. On OOLONG-Pairs, RST with Qwen3-8B achieves an F1 score of 31.7%, outperforming a zero-depth RLM based on the 60× larger Qwen3-Coder-480B (17.3%). With Qwen3-235B, RST reaches 61.6%, compared to 43.9% for a zero-depth GPT-5-based RLM. Comparisons are restricted to zero-depth RLMs because RST itself uses no sub-calls. We also observe substantial improvements on LongBench-v2 CodeQA and consistent gains on standard OOLONG. We further integrated Mamba2 SSM for extraction and compaction within the RLM and RST frameworks, leveraging its efficient long-sequence modeling capabilities. This yielded substantial performance gains, i.e. Qwen3-8B + Mamba2-1.3B in RLM achieved 36% accuracy on OOLONG, matching RLM using GPT-5 (depth 0), while Qwen3-8B + Mamba2 in RST reached 38% accuracy on CodeQA and 53.33% F1 on OOLONG-Pairs, significantly outperforming the corresponding base-model-only RLM and RST variants.