MMLA-RSI: Recursive Self-Improvement through RTT and RTMA
MMLA-RSI proposes recursive self-improvement through two Memory-Mediated Learning Architecture routes: Reasoning-Time Training (RTT) treats the numerical policy-state updater as an object of improvement, while Reasoning-Time Memory Adaptation (RTMA) treats the memory-management rule as an object of improvement. Each admitted successor takes over its parent's role in producing the next generation's rule. The paper's main contribution is a formally bounded construction with shared interfaces and a common algorithm, though the authors explicitly state that beneficial recursive chains remain to be evaluated.
The paper offers a rigorous formal specification of recursive self-improvement but admits that it has not yet demonstrated the central phenomenon. The authors are commendably explicit that 'the two recursive routes and their joint realization are specified here; their beneficial recursive chains and discovery efficiency remain to be evaluated.' The conditional properties---causal composition, finite admission, and finite work---are carefully derived, yet they do not guarantee that an improving successor exists or will be discovered. As a result, the work reads as a precise theoretical scaffold awaiting empirical validation.
The state-account formalism and its three timescales (attempt, problem, improvement generation) are precise and well-structured. The RTT witness obligations in Table 2 rigorously separate process evidence from functional use and task benefit, and the authors rightly insist that 'feedback must arrive before it can teach' with the strict event order $\epsilon_j \leq v_j \leq u_j < \sigma_{j+1} < \tau$. The bounded languages---scalar schedule trees for RTT and complete-row write versus NULL choices for RTMA---give concrete, inspectable domains rather than vague hand-waving.
The fundamental limitation is that the paper specifies recursive self-improvement without exhibiting a single recursive chain. The component studies provide only partial and often negative support: in the CI replay, real feedback binding helped relative to sham ($D = +0.14969$) but retention actually increased loss ($G = -0.05337$), with the authors noting that 'retaining real updates is worse than resetting in every seed.' The SQL studies likewise fail the prespecified online-gate requirements---the adapted CoSQL arm loses to the frozen comparator and the authors report that it 'fails against the frozen method'---while the program-selection task uses only four supplied candidates and is massively outperformed by a direct support-loss rule (0.78125% vs. ~25--30% error). The authors candidly acknowledge that these results 'provide component foundations, rather than a test of the recursive construction's success or failure,' but this leaves the paper's central claim entirely unsupported empirically.
The evidence presented does not support claims of beneficial recursive improvement because no recursive chain is empirically reported. The authors correctly observe that 'a numerical update that helps a task, an updater that improves task learning, and an updater that helps produce a useful successor are observations at different levels,' but the paper provides evidence only for the first level and only in component tasks. Comparisons to related work in Table 4 are fair in distinguishing demonstrated dependencies, while the proposed MMLA-RSI dependency remains purely theoretical. The explicit framing that 'existing component studies motivate the construction' rather than validating it is intellectually honest, but it means the paper's novelty relative to prior routes like STOP, RRSI, and MemEvolve is architectural rather than empirically demonstrated.
Independent reproduction of the proposed recursive system is currently impossible. The authors state clearly: 'No public URL is asserted for an unpublished recursive implementation,' and 'The full private research workspace, raw trajectories, and model weights are not included in this writing package.' While the formal specification includes bounded languages, a trusted committer, and version semantics, the absence of released code, exact hyperparameters for proposal learning (generation cap $K$, proposal cap $N$, sampling temperature), and the missing public implementation block any replication. Component studies reference the LARC and MMLA v4 reports, but without access to the specific bounded-rule interpreter and candidate generation framework described here, the central algorithm cannot be reproduced.
An agent can improve its behavior within an unresolved problem by changing how it acts and which experience it retains. We propose MMLA-RSI, a recursive self-improvement scheme with two implementations in Memory-Mediated Learning Architecture: Reasoning-Time Training (RTT) changes the policy state $\Phi$, while Reasoning-Time Memory Adaptation (RTMA) changes authoritative memory $M$. Each route also makes its adaptation method an object of improvement. A current numerical updater trains the policy proposing its successor; a current memory manager chooses the records available to the process proposing its successor. An independently admitted successor takes over the same role in the next generation. We give bounded executable rule languages, shared task/proposal interfaces, and one algorithm for either route or their composition. The RTT instance changes scalar schedules in a clipped-momentum update. The RTMA instance chooses NULL or a complete-row write with possible eviction. A fixed interpreter, trusted committer, and independent evaluator bound what can change. State versions, read closures, and conditional properties describe causal execution, accepted transitions, and finite work. Existing numerical-component studies motivate the construction. In four-candidate program selection, retaining LARC feedback updates reduces expected error by 24.65 and 36.65 percentage points for two trained initializations. In chronological continuous-integration replay, feedback binding is positive while retention increases loss. Completed SQL studies improve substantially over an unadapted substrate but do not establish the required additional online benefit. These results characterize their declared components. The two recursive routes and their joint realization are specified here; their beneficial recursive chains and discovery efficiency remain to be evaluated.
Pick a starting point or write your own. Challenges run in the background, so you can keep reading while the AI investigates.
No challenges yet. Disagree with the review? Ask the AI to revisit a specific claim.