RSI-Anything: Unlocking Recursive Self-Improvement with World Models

简体中文

This is a concept_demo / L1 mechanism diagnostic, not a reproduction of the original capability evaluation or a candidate for Evolve promotion.

Paper information

Field	Value
Paper	arXiv 2610.13912
Affiliations	PhAI Labs / The Chinese University of Hong Kong and others
First public release	2026-10-09 (arXiv v1)
Author code	Yes: RSI-Anything.
Adapter	rsi-anything
Local implementation	src/auto_research/reproductions/rsi_anything/

Original paper

Background and contribution

A world model predicts candidate outcomes before requesting external verification. Feedback updates a persistent policy workflow. If predictor reliability is poor and enough data and budget remain, adapt the predictor too. Only versions that pass both component and pairwise compatibility checks advance. See the full paper.

Architecture

graph LR
 A[Persistent policy proposes candidates] --> B[Action-conditioned latent prediction]
 B --> C[Select external verification]
 C --> D[Version-attributed feedback]
 D --> E[Update workflow]
 D --> F[Conditionally adapt predictor]
 E --> G[Component and pair gates]
 F --> G
 G --> A

Local redraw; the original Figure 2 shows the paper mechanism and is not replaced by the redraw as evidence. See source provenance.

Key figure from the paper

Original Figure 2: The overall workflow, major stages, and data flow. From the paper; copyright belongs to the authors.

Core equations

\(\widehat z_{t+1}=F_\phi(E_\phi(s_t),a_t,c_d)\) and \(\widehat o=R_\phi(\widehat z)\). Adapt when \(u_k^W=1\{\rho_k^W=\mathrm{unreliable}\}1\{D_{adapt}\text{ is sufficient}\}1\{B_{update}\ge\widehat c\}\). If \(Q_{pair}\) fails, restore the complete previous \((\theta,\phi)\) pair.

Paper results

Metric	Paper results	Protocol or comparison
Code repair: static / predictive / adapted	26.3% / 29.2% / 31.2%	Same coder, approximately matched test usage
Laboratory planning: first / fifth policy	21.80% / 46.01%	Five consecutive policy versions
Best reaction yield	96.15%	Bayesian optimization 82.16%; 40 experiments each

These are paper results, not local measurements. No production A/B was reported. Different models, data, scoring, and resource protocols are not directly comparable. Paper results are kept outside local metrics and the dashboard.

Local reproduction

Compare against the initial workflow and predictor pair. The current implementation executes the experimental mechanism but establishes neither stable relative gain nor a formal capability comparison.

Implementation: src/auto_research/agent_research/oct12_evolution.py and oct12_checkpoint.py. The thin adapter reads explicit oct12.json configuration without fixed default candidates or gold answers.

PYTHONPATH=src python scripts/run_oct12_agents.py \
  --method rsi-anything --checkpoint "$PUBLIC_CHECKPOINT_CACHE" --device cuda \
  --checkpoint-id Qwen/Qwen2.5-7B-Instruct \
  --checkpoint-revision bb46c15ee4bb56c5b63245ef50fd7637234d6f75 \
  --seed 42 --rounds 2 --group-size 2 --train-steps 2 \
  --max-tokens 1536 --output runs/oct12/rsi-anything-7b-1536-seed42/result.json

Run from the repository root in its Python environment. PUBLIC_CHECKPOINT_CACHE must point to a byte-verified local copy of the pinned public revision, not floating main. CUDA requires Linux bubblewrap/prlimit, PyTorch, Transformers, safetensors, and a real A100/A30. Load only local weights with remote code disabled. Do not reuse the output directory. This seed/budget/checkpoint configuration ran on A100; see the receipt for the exact snapshot and sanitized command. The example cache path is not a machine path.

Data and implementation

The input is public-circle-packing-26, without the optimal layout. Training and gate use different context strings but the same underlying task, not independent generalization validation/test. Gate labels do not train the predictor. Short runs set diagnostic_only=true, capability_comparison=false, and evolve_promotion_eligible=false. Generated programs run only under Linux bubblewrap/prlimit; without isolation the runner fails closed instead of executing code on the host.

Validation and limitations

The implementation uses real checkpoint embeddings, a trainable factorized latent predictor, and an outcome readout. Targets encode observed execution feedback using latent MSE, outcome MSE, and factor orthogonality. This is not the authors' JEPA-Anything model. Training and gate labels stay separate. Historical predictions are recomputed under the current version rather than mixing errors across versions.

A real NVIDIA A100 ran Qwen2.5-7B-Instruct, seed 42, 1536 maximum output tokens, 2 rounds, 2 candidates per round, and 2 conditional adaptation steps at snapshot 3eae2c8863678692a8cf2de988d4562314c2731f. Sanitized measurements and round audit are local, not paper results.

Round one reported reliability=insufficient and did not update the predictor. Round two reported unreliable and triggered actual adaptation. Predictor loss was 0.24024838209152222 → 0.19830742478370667, with gradient norms 0.6622512340545654 / 0.5989798307418823. The policy component passed both rounds, the model component only round two, and both pair gates passed. Round-two pair prediction error changed from 0.2382378494734292 to 0.15235200799976756. Policy updates modify persistent workflow, not backbone weights. The predictor is not a JEPA-Anything checkpoint.

Circle validation fitness was zero for old and new pairs in both rounds. Pair retention does compare utility and prediction error, but separate contexts on the same task do not establish independent capability. Lower loss and gate acceptance are not improved circle solving. Keep concept_demo / L1, diagnostic_only=true, capability_comparison=false, and evolve_promotion_eligible=false.

The GPU receipt preserves source hashes from git show at the execution commit, public checkpoint revision, task-spec hash, and actual values. The task downloads no external dataset; environment revision is bound to the source snapshot, not an invented HF revision. Receipt passed applies only to that snapshot's mechanism run. It does not validate later or dirty changes. Keep paper and local metrics separate and retain the all-zero fitness. Raw model outputs, machine paths, logs, and weights are not published.

Tests: python -m pytest tests/test_oct12_agent_algorithms.py tests/test_oct12_agent_delivery.py.