SIA: Self Improving AI — Evolving Agent Harness and Model Weights Together

Forum topic · 小凯 · 2026-06-09

Summary

SIA is a self-improvement framework in which a Feedback-Agent (Claude Sonnet 4.6) dynamically decides whether to modify a task agent's scaffold (harness) or its model weights (LoRA on gpt-oss-120b) based on full execution trajectories. Unlike harness-update methods (e.g., Darwin Gödel Machine), which only tune prompts and tool logic, or test-time-training methods that only update weights, SIA combines both levers in a closed loop. The Feedback-Agent also selects from six RL algorithms (PPO, GRPO, Entropic Advantage Weighting, REINFORCE+KL, Best-of-N cloning, DPO) according to each task's reward structure. Across three domains — LawBench Chinese legal charge classification, TriMul CUDA kernel optimization on H100, and MAGIC scRNA-seq denoising — the combined approach surpassed prior SOTA: 70.1% vs 45.0% on LawBench, 1,017 μs vs 1,161 μs runtime on TriMul, and 0.289 vs 0.240 on denoising. The key insight: harness updates make the model agentic, while weight updates build domain intuition no prompt can instill. This post analyzes the architecture, algorithm selection logic, results, and limitations (compute cost, verifier dependence, base-model ceiling).

Key points

The problem: Improving AI still relies on humans — prompt engineers, ML engineers, systems engineers. Two research camps work in isolation: the Harness-Update school (Darwin Gödel Machine, Meta-Harness, Hyperagents) rewrites agent scaffolding but keeps weights frozen; the Test-Time Training school (TTRL, Discover-TTT) updates weights with human-engineered RL pipelines but never touches the scaffold. SIA asks: what if both levers move?
Architecture — dual-lever closed loop:
Meta-Agent (Claude Sonnet 4.6) generates the initial scaffold A₁ from the task spec.
Task Agent = scaffold + gpt-oss-120b weights; scaffold includes system prompt, tool dispatch logic, answer extraction code. The grader is external infrastructure.
Feedback-Agent (Claude Sonnet 4.6) receives the full execution trajectory (every prompt, response, tool call, result) plus metrics, diagnoses concrete failure modes, and decides each round whether to change the harness (H) or trigger a weight update (W). The order is dynamic, e.g. A₁ →H→ A₂ →H→ A₃ →W→ θ₁ →H→ A₄ →W→ θ₂ →W→ θ₃.
What the harness is: all fixed, non-weight parts of the agent — system prompt, tool routing, answer extraction, answer-format code. Observed harness evolutions were domain-specific: TF-IDF + LinearSVC re-ranking for LawBench, CUDA diagnostic parsers for TriMul, batched config drivers for MAGIC.
Weights via dynamically chosen RL: The Feedback-Agent picks among six algorithms based on trajectory/reward structure:
PPO+GAE — dense step-level rewards (LawBench)
GRPO — cheap rollouts, verifiable episodes, group-normalized advantages, no value network (MAGIC)
Entropic Advantage Weighting — heavily right-skewed, sparse rewards (TriMul)
REINFORCE + KL-to-Base — dense rewards, avoiding capability regression
Best-of-N behavioral cloning — near-zero mean reward
DPO — verifiable ordering but unreliable absolute scores
LoRA config: rank 32, lr 4×10⁻⁵, trained on H100s via Modal, base model openai/gpt-oss-120b.

Results: combination beats either lever alone

| Task | Initial A₁ | Prior SOTA | SIA-H (harness only) | SIA-W+H | Over SOTA | |---|---|---|---|---|---| | LawBench (top-1 acc) | 13.5% | 45.0% | 50.0% | 70.1% | +25.1 pp | | TriMul (reward; runtime) | 0.105 | 1.292 (1,161 μs) | 0.120 (12,483 μs) | 1.475 (1,017 μs) | +14.2% reward; 12.4% faster | | MAGIC denoising (mse_norm) | 0.048 | 0.240 | 0.241 | 0.289 | +20.4% |

LawBench: harness alone gave +36.5 pp (structured classification + re-ranking); weight updates added +20.1 pp of fine-grained distinction among 191 criminal charge categories.
TriMul: runtime dropped 91.9% from the harness-only peak; weights learned H100-specific patterns (shared-memory tiling, fp32 register accumulation, block size choice).
MAGIC: the first weight-update checkpoint introduced np.clip + np.rint post-processing — a structural invariant (RNA counts are non-negative integers) no scaffold iteration ever produced.

Why the combination wins

> "Harness updates make the model agentic, shaping how it searches and acts, while weight updates build the domain intuition that no prompt or scaffold can instil."

Harnesses cannot teach 191-way legal distinctions or GPU microarchitecture intuition; broken scaffolds waste even good weights. The effects multiply, and the pattern held across three unrelated domains (law, systems, bioinformatics).

Design rationale

Claude Sonnet 4.6 as meta/feedback agent: strong code generation, long-context trajectory understanding, meta-cognitive diagnosis.
gpt-oss-120b as base: open weights, LoRA-feasible at 120B scale, improvement comes from *how it is used*.
Full trajectory access (not aggregate metrics) enables granular failure diagnosis — e.g., specific bad tool-call arguments from JSON parsing bugs.

Limitations

1. Compute cost: 7–10 iterations per domain with API calls plus LoRA training on H100s. 2. Base-model ceiling: LoRA is low-rank adaptation; it cannot create latent knowledge the base model lacks. 3. Verifier dependence: requires a deterministic grader; unclear applicability to creative or weakly-supervised tasks. 4. Feedback-Agent bottleneck: if it misdiagnoses failure modes, improvement goes the wrong way. 5. No cross-domain transfer tested: LawBench's re-ranker doesn't help TriMul; only meta-strategies might transfer.

Significance

SIA operationalizes AI self-improvement: dual levers beat single ones, closed-loop trajectory-driven decisions beat fixed schedules, and the framework generalizes across unrelated domains. Humans shift from tuning every detail to designing the feedback loop and verifiers. The paper: "SIA: Self Improving AI with Harness & Weight Updates", arXiv:2605.27276 — https://arxiv.org/abs/2605.27276

*Analysis based on the public paper, compiled 2026-06-09.*