Harness-Aware Self-Evolving (HASE) Framework

• HASE is an agentic reinforcement-learning framework that unifies task solution generation and harness editing in a multi-turn process.
• It overcomes fixed-harness limitations by allowing selective, safe modifications of guidance and evaluation components based on real-world feedback.
• Experimental results demonstrate that HASE enables small models to match or surpass benchmarks in text classification, alpha factor mining, and algorithm discovery.

Harness-Aware Self-Evolving (HASE) is an agentic reinforcement-learning framework in which a single model can generate task solutions or edit selected harness components in a multi-turn action space, so that the harness and the solution improve through one unified process rather than under the usual fixed-harness assumption (Luo et al., 4 Jul 2026). The framework is motivated by two limitations of fixed-harness self-evolution: evaluation misalignment, in which human-written evaluators may be incomplete or misaligned with external reality, and guidance deficiency, in which prompts, memory, retrieval, or other surrounding machinery remain suboptimal even when evaluation is correct. In the reported experiments, HASE enables a single Qwen3-8B model to match the text-classification performance of a GPT-OSS-120B model that uses Claude Code as the harness proposer, outperforms the reported GPT-OSS-120B baseline in alpha factor mining, repairs imperfect evaluation components, and converges to state-of-the-art performance in circle-packing algorithm discovery (Luo et al., 4 Jul 2026).

1. Conceptual setting and scope

Self-evolving systems have typically optimized task solutions while treating the surrounding harness as fixed. HASE targets precisely that assumption. In the surrounding literature, the harness is described as the runtime substrate or runtime interface that mediates how a model observes its environment, acts, receives feedback, and is constrained by prompts, tools, memory, orchestration, and control flow (Zhong et al., 13 May 2026). HASE makes parts of that substrate editable from within the learning loop itself, but only on a selected and whitelisted surface (Luo et al., 4 Jul 2026).

Within the broader harness literature, this places HASE between two previously separated traditions. One tradition treats harness engineering as the main lever, emphasizing prompts, tools, memory, middleware, or workflow search; another focuses on reinforcement-learning adaptation while leaving the surrounding scaffold fixed. HASE does not separate these into distinct systems. Instead, one model alternates between object-level task behavior and harness-level intervention inside a shared trajectory, thereby co-evolving policy, harness, and solution (Luo et al., 4 Jul 2026).

This design also differs from frameworks in which a separate or stronger external agent proposes harness changes. HASE explicitly states that there is no external proposal LLM: harness editing and solution generation are unified within the same model, with only a deterministic reviewer at phase boundaries (Luo et al., 4 Jul 2026). A plausible implication is that HASE treats harness improvement not as offline engineering but as an endogenous part of agent behavior.

2. Editable harness surfaces

HASE distinguishes between guidance components and evaluation components. Guidance components include prompts, retrieval, context formatting, memory, and related machinery that improve search or execution. Evaluation components are local evaluators that judge solution validity. The distinction is operational rather than merely descriptive: guidance components can be freely modified to improve search and execution, whereas evaluation components can be edited only under strict real-world feedback anchoring to avoid reward hacking (Luo et al., 4 Jul 2026).

The editable surface is task-specific and explicitly whitelisted.

Task	Editable harness components	Type
Heilbronn (triangles)	Prompt + evaluator	Eval + Guidance
Circle packing	Prompt + evaluator	Eval + Guidance
Text classification	Context + memory	Guidance
Alpha mining	Context + prompt	Guidance

This whitelisting is central to the framework’s safety envelope. HASE does not claim unrestricted self-modification of all runtime infrastructure. The reported limitation is a conservative surface of editable components, and broader application is left open (Luo et al., 4 Jul 2026).

The distinction between guidance and evaluation also clarifies a common misunderstanding. HASE is not a generic mechanism for rewriting the score function whenever the agent benefits. Evaluation edits are permitted only when the real-world evaluator disagrees with the local proxy, and acceptance depends on repairing those mismatches rather than merely increasing proxy reward (Luo et al., 4 Jul 2026).

3. Unified multi-turn reinforcement-learning loop

The core methodological innovation is a unified, multi-turn action space. At each turn, the model can submit an object-level solution, inspect or modify harness files via a sandbox, or terminate the interaction with done. The sandbox actions listed in the description are ls, cat, grep, replace, and overwrite. There is no hard-coded planner: the model chooses whether to solve the task directly or intervene on the harness based on the current trajectory context (Luo et al., 4 Jul 2026).

Training proceeds in phases, each containing multiple RL epochs such as GRPO or PPO. During rollout, harness edits are trajectory-local; only at phase boundaries are persistent harness changes reviewed and possibly committed. The formal loop is summarized as follows:

1. Initialize policy $\pi_\theta$, harness $h_0$, and local evaluator $E_{\mathrm{loc}}$.
2. For each phase, roll out multi-turn trajectories $\tau \sim \pi_\theta(\cdot \mid x, h)$.
3. Score solutions by proxy with $E_{\mathrm{loc}}$.
4. Collect candidate guidance edits whose edited-harness trajectories demonstrate improved task performance.
5. If evaluator edits are allowed, query the real-world evaluator, add disagreements to the mismatch set $\mathcal{M}$, and review evaluator patches for mismatch improvement.
6. Update the policy by policy gradients.
7. Commit the best stable harness at the phase boundary (Luo et al., 4 Jul 2026).

For guidance edits, HASE uses a proof-of-concept shortlisting signal:

$$R_{\text{PoC}}(\tau) = \max\{0, R_{\text{task}}(\tau; h') - R_{\text{base}}(x, h)\}.$$

Only edits that produce a higher downstream score in at least one trajectory are shortlisted and considered at the phase boundary (Luo et al., 4 Jul 2026).

A key consequence is the separation between transient experimentation and persistent harness evolution. Trajectory-local edits can influence a single rollout, but only phase-committed edits become part of future episodes. This prevents a local exploratory change from immediately corrupting persistent state.

4. Co-evolution of policy, solutions, and evaluators

HASE jointly co-evolves three objects: model policy or weights, task solutions, and harness components. Policy updates occur through agentic RL after each trajectory. Solutions evolve through iterative proposal and phase-review selection. Harness components evolve either as guidance edits or as evaluator repairs, depending on the task and the observed mismatch pattern (Luo et al., 4 Jul 2026).

The evaluator-repair mechanism is the most distinctive part of the framework. HASE begins with the possibility that a local evaluator is incomplete and therefore exploitable. It maintains a mismatch set $\mathcal{M}$ of proxy–oracle disagreements. Evaluator patches are committed only if they repair oracle–proxy mismatches, maintain the interface, and pass non-regression checks. This directly addresses reward hacking not by prohibiting evaluator edits altogether, but by constraining them to realized discrepancies against external feedback (Luo et al., 4 Jul 2026).

The circle-packing and Heilbronn settings exemplify this design. In circle packing, HASE starts from an incomplete evaluator that omits key constraints. The system detects proxy–oracle divergences, goes through two repair cycles, and aligns proxy and oracle scores after repair. In Heilbronn triangles, the initial evaluator lacks boundary checks, the model exploits that omission by placing points outside the triangle, the oracle gives zero, and HASE synthesizes and commits boundary checks to close the exploit gap (Luo et al., 4 Jul 2026).

These experiments support a broader claim about co-evolution. Freezing policy after evaluator repair leads to a plateau, whereas continuing RL with the repaired harness allows the agent to internalize the corrected constraints and improve further. In the Heilbronn setting, continuing RL after repair reaches $0.036082$ compared with $0.021880$ when the policy is frozen after repair (Luo et al., 4 Jul 2026). This suggests that harness correction and policy adaptation are complementary rather than interchangeable.

5. Experimental results

HASE is evaluated on four benchmark families: text classification, alpha factor mining, circle-packing algorithm discovery, and Heilbronn triangles. The reported results emphasize both small-model competitiveness and evaluator repair (Luo et al., 4 Jul 2026).

Domain	HASE result	Comparator
Text classification (Symptom2Disease)	$h_0$0 test accuracy	GPT-OSS-120B + Meta-Harness: $h_0$1
Alpha factor mining	RankIC $h_0$2, AER $h_0$3, IR $h_0$4	GPT-OSS-120B: RankIC $h_0$5, AER $h_0$6, IR $h_0$7
Circle packing	Packing score $h_0$8	AlphaEvolve best-known result $h_0$9
Heilbronn triangles	$E_{\mathrm{loc}}$0 with continued RL after repair	$E_{\mathrm{loc}}$1 frozen after repair

In text classification, the editable harness components are context retrieval and memory formatting, and ground-truth evaluation is not edited. Under that setup, Qwen3-8B + HASE-harness achieves $E_{\mathrm{loc}}$2 test accuracy, matching or slightly surpassing the much larger GPT-OSS-120B + Meta-Harness system at $E_{\mathrm{loc}}$3 (Luo et al., 4 Jul 2026). The reported qualitative outcome is that the agent learns to overwrite retrieval scripts for semantic or symptom-based search, repair error-prone tools, and build memory summaries for label disambiguation.

In alpha factor mining, HASE evolves prompt and pool-context presentation while keeping evaluation immutable. The reported Qwen3-8B result is RankIC $E_{\mathrm{loc}}$4, AER $E_{\mathrm{loc}}$5, and IR $E_{\mathrm{loc}}$6, compared with GPT-OSS-120B at RankIC $E_{\mathrm{loc}}$7, AER $E_{\mathrm{loc}}$8, and IR $E_{\mathrm{loc}}$9 (Luo et al., 4 Jul 2026). The harness evolution is described as a progression from “recency-only” to IC-ranked retrieval, driven by evolved context-engineering scripts and prompt refinements.

In circle packing, the significance lies less in scale comparison than in evaluator repair: HASE reaches a packing score of $\tau \sim \pi_\theta(\cdot \mid x, h)$0, matching AlphaEvolve’s best-known result after repairing missing boundary and overlap checks (Luo et al., 4 Jul 2026). In Heilbronn triangles, the principal result is the demonstration that repaired evaluation and continued RL outperform repaired evaluation with frozen policy.

Taken together, these results suggest that a substantial part of observed capability can reside in how a model edits and exploits its runtime substrate, not only in parametric scale. The paper itself makes the narrower empirical claim that harness evolution enables small models to reach frontier performance on harness-sensitive tasks (Luo et al., 4 Jul 2026).

6. Relation to the wider harness-evolution literature

HASE belongs to a rapidly expanding research area that treats the harness as a first-class optimization object. That literature includes fully automated harness engineering loops, trace-based observability frameworks, typed and composable harness foundries, regression-gated self-editing, retrospective optimization from unlabeled trajectories, and diagnosis-driven repair operators (Seong et al., 22 Apr 2026). HASE’s distinguishing feature within that landscape is the unification of solution generation and harness editing inside one model and one multi-turn RL process, rather than delegating harness evolution to a separate meta-agent or a stronger proposer model (Luo et al., 4 Jul 2026).

Nearby systems illuminate the design space. SIA combines harness updates with model weight updates through a distinct Feedback-Agent and reports strong gains from interleaving those two levers across legal classification, GPU kernel optimization, and scRNA-seq denoising (Hebbar et al., 26 May 2026). Self-Harness likewise removes human engineers and stronger external agents, but organizes improvement into Weakness Mining, Harness Proposal, and Proposal Validation over a minimal initial harness (Zhang et al., 8 Jun 2026). Trace-driven systems such as AHE and HarnessFix emphasize observability, attribution, and scoped repair, arguing that stable self-evolution depends on explicit componentization, distilled trajectory evidence, and regression-aware validation (Lin et al., 28 Apr 2026).

The broader literature also introduces cautionary results. “Harness Updating Is Not Harness Benefit” separates the capability to produce useful harness updates from the capability to benefit from them, finding that weak-tier models may fail to activate relevant harness artifacts or fail to follow them faithfully (Lin et al., 28 May 2026). SEAGym, an evaluation environment for self-evolving agents, reports that frequent updates may fail to improve held-out performance, useful intermediate snapshots may collapse later, and model backend and source diversity affect harness reliability (Zheng et al., 16 Jun 2026). These findings do not invalidate HASE’s reported results, but they indicate that harness-aware self-evolution should be evaluated not only by a single improvement curve, but also by transfer, regression, replay, and cost.

HASE’s own limitations are stated narrowly: the editable surface is conservative, and scaling is non-exhaustive (Luo et al., 4 Jul 2026). A plausible implication is that future work will depend on how far editable harness surfaces can be widened while preserving the anchoring and non-regression properties that make evaluator repair and guidance adaptation empirically tractable.