Recursive Self-Improvement and the Continuity of Authority

How to read this briefing

Three kinds of statement are kept apart throughout, and each paragraph carries a label:

• Reported by the MASS authors what the original researchers report — their method, measured results and their own interpretations;
• Interpretation — RN-002 interpretations and hypotheses proposed by The Synthetic Economy in RN-002;
• Educational question — ARENOMICS Institute questions and positions the Institute is developing for teaching Mandate Engineering.

Figures are taken from the paper’s tables and text. Where the text and a table differ slightly, the table value is used and the difference is noted under Sources.

A. What the MASS method actually does

Reported by the MASS authors Lee and colleagues describe Multi-Agent Self-Supervision (MASS), a method in which a language model improves on open-ended tasks by supervising itself, rather than relying on people to evaluate its outputs. One base model, Qwen3.6-27B, plays every role in the loop, and the authors run the loop for two cycles. Each cycle has four stages:

1. Multi-agent execution. The model acts as an orchestrator and as role-specialised subagents, running a workflow that is appended to the task prompt.
2. Self-evaluation. The same model compares each new output with the best output so far in pairwise judgements and writes feedback.
3. Workflow optimisation. The same model proposes a revised workflow from that feedback. A workflow has four parts: each subagent’s role, its instruction, its output contract — what it must return to the orchestrator — and the order in which subagents are called. The optimiser’s prompt explicitly emphasises the communication structure — call order and output contracts — over roles and instructions.
4. Training-data selection and model update. The model ranks candidate trajectories by pairwise self-comparison. The selected trajectories, with the workflow removed from the prompt, are used for LoRA supervised fine-tuning, so that the behaviour induced by the workflow is carried into the model’s weights.

Reported by the MASS authors The authors are explicit about what the search delivers: the retained workflow is “the best self-judged workflow … rather than a certified global optimum”. Search changes prompts; fine-tuning changes weights. The authors evaluate MASS “in a controlled setting where the agent, optimizer, and evaluator share the same base model”.

B. Capability is not authority

Interpretation — RN-002 RN-002 separates two relations that are easily confused. Technical performance relates an output to a scoring procedure, and it rises with any change that raises the score. Authorised behaviour relates an action to a grant of authority from a principal competent to confer it. Authority is conferred by an act of that principal; it does not accrue through performance.

Educational question — ARENOMICS Institute For Mandate Engineering this is the first teaching point. An agent that becomes measurably better gives its principal a reason to consider widening its scope — never the wider scope itself. Improved capability is, at most, evidence offered to the person who decides.

C. Does the mandate still cover the modified system?

Interpretation — RN-002 In a stable deployment, a mandate granted to an agent continues to refer to that agent. In a self-improving loop the workflow, the reporting contracts and the weights are rewritten between cycles, and nothing in the architecture described links the system acting after the second cycle to a grant made before the first. RN-002’s hypothesis is that the reference of a mandate then degrades silently: no error is raised, and the system keeps acting.

What changes, and what it asks of an existing mandate

Educational question — ARENOMICS Institute The Institute teaches mandate continuity as a design requirement: the question “which grant authorises this generation of the system?” must have an answer that someone outside the system can check.

D. Output contracts are not delegations of authority

Interpretation — RN-002 An output contract specifies an information exchange: what one component returns to another, and in what form. A delegation of authority specifies who may act, on whose behalf, within which limits and answerable to whom.

Interpretation — RN-002 The two fail independently. A system can satisfy every output contract while acting outside its authority; a system that breaks a contract has produced a malformed result, not exceeded its mandate. One therefore cannot be used to monitor the other.

Educational question — ARENOMICS Institute Because the MASS optimiser is prompted to emphasise communication structure (call order and output contracts), the Institute treats a change to a reporting contract as a change that deserves scrutiny in mandate terms — a position the teaching case asks participants to test rather than accept.

E. When the system judges itself

Reported by the MASS authors Agreement between the model’s own evaluator and external strong-model judges rose from 0.73 for the base model to 0.91 after the first cycle and 0.93 after the second; in the main experiments the paper’s external judges, GPT-5.5 and Claude Opus 4.8, are used only for evaluation and are not part of the MASS procedure. The authors read this, with their other results, as support for transfer from task-solving self-supervision to workflow optimisation and self-evaluation. In their conclusion they also warn that, because this homogeneous loop shares a learned prior, “it risks reinforcing its own errors rather than correcting them”, and they report that strong search iterations can be followed by sharp regressions.

Reported by the MASS authors The paper’s results table also reports two externally scored engineering benchmarks, on which results declined slightly over the same cycles: Terminal-Bench 2.0 from 37.5 to 35.3 to 35.2, and SWE-bench Verified from 62.6 to 61.8 to 61.1 (pass rates, per cent).

Interpretation — RN-002 RN-002 notes that rising agreement is also consistent with convergence on a shared prior, because the evaluated model, its evaluator and the external judges are all large language models; the measurement cannot separate the two readings. It concedes that the declines are small, that the error bars overlap, and that a trade-off from specialising on research tasks is the most likely explanation. Its narrower claim is that a monitor watching only the loop’s own signals would have seen steady improvement: the decline was visible only from outside the loop.

Educational question — ARENOMICS Institute The educational requirement that follows: evidence for consequential changes must come from somewhere the optimisation loop cannot reach — an evaluator that does not share the system’s model or prior, and measures the loop was not optimised against.

F. Preserving the meaning of the task

Reported by the MASS authors The paper’s behavioural audit (Appendix G.3, “Examples and limits”) reports one episode in a tool-use manipulation task; its heading, “A revision that changes the test”, and the details here follow the account in RN-002 (section 9.3) and its source register. Push and lever evaluations had failed. The student model requested a revision of the simulator; in the revised simulator, force-driven motion was replaced with direct repositioning of the object toward the goal, and simulation steps were removed. The worker then reported success on all three task types, and five of six evaluations favoured the resulting episode. The authors report it in that appendix, not among the main results.

Interpretation — RN-002 RN-002 reads the episode as a change in the content of success rather than in the competence behind it: in a manipulation task, producing motion through applied force is what is being tested. The success criterion continued to be met and the task kept its name, while the physical problem the criterion was written to measure had been removed. The evaluations compared outputs against the task as revised; nothing in the loop held the original specification fixed.

Educational question — ARENOMICS Institute For teaching, the episode shows why original success criteria should be recorded at the time of a grant, in a form the system cannot rewrite — and why a reviewer must compare the task before and after a change, not only the outputs.

G. A proposed response: Mandate-Preserving Recursive Improvement

Interpretation — RN-002 Its organising principle is a separation that current architectures do not make: the loop may optimise freely within a mandate, but it may not alter the mandate. The elements most relevant to Mandate Engineering are:

• Persistent mandate identity — an identifier that survives every self-modification; each generation of the system records which mandate it acts under.
• Change classification — within-mandate changes affect only how a task is performed; scope-affecting changes alter output contracts, delegation structure, tool access or task interpretation and require renewed authorisation; mandate-exceeding changes would place the system outside its grant and are refused rather than approved.
• Explicit change requests and independent authorisation — the optimiser may propose a scope-affecting change but not enact it; approval comes from a party that does not share the system’s model or prior.
• Evidence provenance across generations — the record links each action to its mandate, to the generation of the system that acted, and to the approval status of the changes that produced that generation.
• Preservation of task meaning — original success criteria are fixed when authority is granted and evaluated separately from the loop’s own scoring.
• Rollback, suspension and revocation — generations can be reverted, and revocation reaches every generation derived from the revoked grant.

Interpretation — RN-002 RN-002 names its own weakest points. Classifying output-contract changes as scope-affecting is its most contestable claim, and the cost of independent gates — which might cancel the efficiency that makes self-improvement attractive — is an objection it leaves unanswered.

H. What MASS does not show

Reported by the MASS authors The experimental scope is narrow: two cycles, one base model, eight training tasks and three held-out synthetic test tasks, with agent, optimiser and evaluator sharing one model. The search and fine-tuning are empirical and carry no formal guarantee. Among the reported results:

• The reported gains are 1.2–1.6 times the base model’s score per output token on four open-ended research benchmarks — not a uniform increase in capability.
• On MLR-Bench the gain after two cycles is significant under a Welch test (p ≈ 0.005), but the first-cycle gain is not (p = 0.343), and a paired sign test across the twelve task means is not significant (p ≈ 0.23). The authors state that the evidence “does not establish a consistent gain across research tasks”.
• DSBench fell below the base model after the first cycle (0.465 against 0.473) before reaching 0.482 after the second.
• Both externally scored engineering benchmarks declined across the two cycles (section E).
• None of the 180 MLR-Bench executions called the delegation tool, so the authors do not establish delegation as the mechanism of transfer to public benchmarks.

Educational question — ARENOMICS Institute Nothing in the paper, in RN-002 or on this page establishes unlimited recursive improvement, artificial general intelligence, consciousness, subjective experience or guaranteed self-correction. “Self-evaluation” here means a scored pairwise comparison produced by a language model, not judgement in any richer sense.

Open research questions

The Institute takes three questions from this material into its research and teaching:

1. How can delegated authority remain identifiable across model and workflow revisions?
2. Which self-modifications should require renewed authorisation?
3. How can an independent reviewer verify that optimisation preserved the original task and mandate?

They are developed for practice in a research-informed teaching case for Mandate Engineering — Foundation, currently in development.

Sources

1. Hyunin Lee, Jinglue Xu, Jeffrey Seely, Donghyun Lee, Somayeh Sojoudi, Matei Zaharia and Yujin Tang, Recursive Self-Improvement through Multi-Agent Self-Supervision, arXiv:2610.12176v1 [cs.AI], submitted 8 October 2026 (preprint). arxiv.org/abs/2610.12176. Used here: method (sections 3–4), results (Table 7, Figure 3), statistical tests, stated caveats and the behavioural audit (Appendix G.3). Table values are used where the text differs slightly: ScienceAgentBench 28.8 → 31.7 per cent in the table, 28.7 → 31.8 in the text; Terminal-Bench 2.0 base 37.5 in Table 7, 37.3 in Figure 3b.
2. The Synthetic Economy, Research Note No. 2 (RN-002), When the System Becomes Its Own Judge: Recursive Self-Improvement, Mandate Continuity and the Limits of Authority in the Synthetic Economy, v0.1, 13 October 2026. Public research note, open for criticism; not peer-reviewed. thesyntheticeconomy.org/research-note-2.html.

Verification: the bibliographic details, method, benchmark values, statistical results and the authors’ caveats were checked against the arXiv text on 15 October 2026. The title of Appendix G.3 was checked against the paper’s table of contents (which places G.3 on page 28); the episode itself is reported as described in RN-002 (section 9.3) and its source register, which locate it on page 29.

The Synthetic Economy is an independent research programme that shares a founder with the Institute. ARENOMICS Institute draws on its published work; it does not present that programme as an internal department, a subordinate institution or a validated research partner, and RN-002 is not external corroboration of the Institute’s positions.