Skip to content

Knowledge-centric self-improvement as a claim economy

Source: Knowledge-Centric Self-Improvement (arXiv:2607.19592v1, 21 Jul 2026) — Xuefei Julie Wang, Lauren Hyoseo Yoon, Chengrui Qu, Amanda Zichang Wang, Atharva Sehgal, Eric Mazumdar, Yisong Yue (Caltech). Code: recursive-knowledge/KSI. Studied 2026-08-06.

Companion to erdos-ai-ledger-lessons.md: that doc reads a human-curated ledger (Tao's Erdős wiki) as a claim economy. KSI is the fully-automated version of the same object — a shared knowledge base where agents emit claims, peers corroborate or challenge them, and a distillation step decides what survives into the next generation. It is the closest published system to what SWARM's claim-ledger scenarios simulate, and it ships with benchmark results.

What the paper does

Invert what persists across a self-improvement run. Agent-centric systems evolve the agent (prompts, workflows, harness, or self-modifying code); KSI keeps agents generic, stateless, and disposable and makes a curated knowledge base the only object that changes. Each agent gets a fresh context, a distilled bundle, and one task attempt, then writes evidence back.

Curation runs in three stages, structured as threaded discussion:

  1. Task-level forum — after attempting a task, each agent posts evidence-grounded claims to that task's thread (what worked, what failed, which constraints mattered), citing and replying to earlier posts.
  2. Cross-task forum — agents propose claims that recur across tasks. Each cross-task claim must be grounded in concrete evidence from one or more attempts, and later posts take an explicit stance: agree, disagree, or synthesize.
  3. Distillation — surviving claims are consolidated into per-task and cross-task bundles. The distiller is instructed to keep claims that are actionable, evidence-grounded, and scoped, and to drop vague advice that does not name the condition under which it applies.

The authors explicitly treat disagreement as evidence rather than instability: conflicting claims expose which guidance is task-specific and which principles hold under contrary evidence.

Results

Haiku 4.5, 50-task subsets, 10 generations, 3 seeds (baselines single-run):

Benchmark KSI Best agent-centric baseline
ARC-AGI-1 86.7% ±4.2 @ $76 HyperAgents 70% @ $234
ARC-AGI-2 82.7% ±6.1 @ $80 HyperAgents 60% @ $188
Polyglot 68.0% ±2.0 @ $126 DGM 58% @ $281
SWE-bench Pro 64.0% ±2.0 @ $208 DGM 54% @ $713
Terminal-Bench 2 43.8% ±3.4 Meta-Harness 37.6%

Higher solve rate at roughly a third of the cost. Against prompt optimization the margin is wider (ARC-AGI-1: 86.7 vs GEPA 44, OpenEvolve 54). The protocol is not LLM-specific — GPT-5.4-mini improves every cell (ARC-AGI-1 93.3% @ $16). Frozen bundles transferred zero-shot to 20 held-out hard tasks help in every donor→recipient pairing, including across LLM families.

Caveats worth carrying

  • The transfer result is the paper's most interesting claim on its thinnest evidence: 20 tasks, and the Haiku→GPT ARC-AGI-1 cell is 38.3% ±12.6. The authors say so — "indicative rather than precise."
  • Terminal-Bench baselines are compiled from the Meta-Harness paper rather than rerun, and one pairs an Opus 4.6 proposer with a Haiku 4.5 base. The authors flag this.
  • Whether the reported dollar cost includes forum and distillation overhead or only task execution is not stated in the main text. The cost story is load-bearing for the paper's Pareto-frontier claim.

The core observation

KSI removes exactly the substrate SWARM's governance levers act on.

Every mechanism in swarm/governance/staking, reputation, random audits, circuit breakers — binds an agent to the consequences of its claims across time. KSI's central design choice is that agents are generic, stateless, and disposable: they carry no persistent identity, no private memory, no reputation. The paper treats this as the source of its benefits (improvements can't overfit to one agent), and for capability that argument holds.

But it means the knowledge base has no reputational collateral behind any claim. An agent that posts a poisoned claim is gone by the next generation, and there is nothing to slash, downweight, or ban. The only thing standing between a claim and the distilled bundle is peer corroboration in the cross-task forum — and corroboration is the cheapest signal in the system to manufacture. This is precisely the failure mode bead 77ao (corroboration-weighted acceptance with collusion discount) was filed to measure, and KSI is a live, benchmarked instance of the vulnerable design.

The second observation follows from the paper's own headline result. Transferability is presented purely as an asset: bundles curated under one LLM family remain effective under another. Symmetrically, a poisoned claim inherits that same portability. The distilled bundle is the only persistent object, so it is also the entire attack surface, and the transfer experiment is evidence that a successful poisoning would propagate across LLM families rather than staying contained to the run that produced it. A single corrupted distilled claim reaches every future agent by construction.

Mapping to SWARM constructs

KSI mechanism SWARM construct Attack it invites
Task-level forum posts Claims of varying p entering a ledger (bead a2il) Volume flooding: cheap plausible claims crowd the verification budget
Cross-task forum: claims survive by recurring across tasks, peers agree/disagree/synthesize Corroboration-weighted acceptance (bead 77ao) Manufactured convergence — a coalition posts mutually-supporting claims across tasks to simulate independent recurrence
Distillation keeps "surviving" claims Acceptance gate with scarce verification Adverse selection: quality_gap < 0 if agreeable-but-wrong claims outlive contested-but-right ones
Only distilled claims reach the next generation; attempts are discarded Survivorship gap E[p\|accepted] − E[p\|attempted] (bead 81sk) Wins-only curation overstates the bundle's quality
Claims cite earlier claims Claim-dependency contamination (bead oa5h) Falsification cascade through citations of accepted-unverified claims
Generic, stateless, disposable agents No analoguestaking/reputation/audit all assume persistent identity No collateral to slash; governance levers are disabled by design

The last row is the research gap. SWARM's existing levers cannot be applied to KSI as specified, because they all price an agent's future. Either the paradigm needs a claim-side substitute for reputational stake (bond the claim rather than the agent), or knowledge-centric self-improvement is structurally more poisonable than agent-centric self-improvement — which would be a real safety cost against a real capability gain, and exactly the kind of tradeoff SWARM exists to quantify.

What this licenses

The KSI protocol is fully specified and the code is public, so the a2il claim-ledger scenario now has an external referent to calibrate against rather than being purely synthetic. Two concrete follow-ups are filed as beads: sx6c (a KSI-shaped acceptance scenario, depends on a2il) and cx83 (a claim-bond lever that restores stake without restoring persistent agents, depends on sx6c).

Open question the paper does not ask, and SWARM can: at equal verification budget, does a corroboration-weighted gate over disposable agents admit more manufactured convergence than a deep single review over identified agents? That is 77ao with KSI's parameters plugged in.