VACE: Validation-Gated Alternating Co-Evolution of Agent Models and Harnesses

Jiexing Qi; Yu He; Jun Liu; Qichen Huang; Shaohua Hu; Zhan Dang; Guohua Chen; Rui Yang

cs.LGSep 29, 2026

VACE: Validation-Gated Alternating Co-Evolution of Agent Models and Harnesses

Authors: Jiexing Qi, Yu He, Jun Liu, Qichen Huang, Shaohua Hu, Zhan Dang, Guohua Chen, Rui Yang, +4 more

Organizations: ICT AI Competence Center, Huawei Technologies Co., Ltd., Shanghai, China · Shanghai Jiao Tong University, Shanghai, China

Abstract

Language model agents can be improved by updating their model weights or refining the harness that guides task execution. These components are coupled: weight updates change how the model uses the harness, while harness updates change the trajectories used for training. We propose VACE, Validation-Gated Alternating CoEvolution, which alternates agentic reinforcement learning with trajectory-driven harness refinement. After each RL stage, VACE reuses the collected trajectories to propose a harness revision and evaluates the incumbent and candidate with the updated model held fixed. The candidate guides subsequent training only if it improves validation performance. With Qwen3.5-9B, VACE achieves 45.26% test accuracy on OfficeQA and a mean partial-credit score of 75.19% on AutomationBench, exceeding weight-only RL by 6.43 and 9.09 percentage points and ungated alternation by 4.59 and 6.95 points, respectively. Across 44 harness proposals, 17 reduce validation performance at the updated checkpoint and are rejected before subsequent RL training, highlighting the importance of validation gating.

Figures & tables

Figure 1: Conceptual schedules for model–harness co-optimization. Harness-only and weight-only update one component. SIA-style and WHALE-style illustrate stagewise and alternating schedules, respectively; VACE validates proposed harness updates before the next training round. The figure uses $\theta,h$ for the weights $W$ and harness $H$ in our notation. The paths and contours are schematic, and the star marks a conceptual joint optimum.

1:	for $t=0,\ldots,T-1$ do
2:	$(W_{t+1},\mathcal{T}_{t})\leftarrow\mathcal{M}(W_{t};H_{t},\mathcal{D}_{\mathrm{train}})$
3:	$H^{\prime}_{t}\leftarrow\mathcal{S}(H_{t},\mathcal{T}_{t})$
4:	$b_{t}\leftarrow\widehat{R}_{\mathcal{V}}(W_{t+1},H_{t})$
	$c_{t}\leftarrow\widehat{R}_{\mathcal{V}}(W_{t+1},H^{\prime}_{t})$
5:	if $c_{t}>b_{t}$ then
6:	$H_{t+1}\leftarrow H^{\prime}_{t}$

Algorithm 1 VACE: alternating updates with validation-gated harness acceptance

Split	HR	Marketing	Finance	Total
Train	26	27	27	80
Validation	15	22	21	58
Test	24	23	24	71
Total	65	72	72	209

Table 1: AutomationBench task counts by domain and split used in our experiments.

Method	OfficeQA	AutomationBench
Method	OfficeQA	HR	Marketing	Finance	Overall
Static agent	$32.11$	$43.86$	$52.57$	$54.44$	$50.26$
Harness-only	$35.28$	$55.03$	$64.53$	$65.42$	$61.62$
RL-only	$38.83$	$63.98$	$67.55$	$66.85$	$66.10$
SIA-style [ Hebbar et al., 2026 ]	$40.67$	$56.36$	$68.83$	$71.01$	$65.35$
WHALE-style [ Kim et al., 2026 ]	$40.67$	$61.50$	$\mathbf{72.30}$	$71.11$	$68.25$

Table 2: Mean test scores (%) from three evaluation runs of each selected model–harness pair: accuracy on OfficeQA and mean partial credit on AutomationBench. Bold marks the highest listed mean in each column.

Dataset	Initial	Final retained	Best	H trials	A/T/R
OfficeQA	30.28	47.17	60.37	12	7/2/3
AutomationBench	53.08	81.42	81.42	32	18/0/14

Table 3: Phase-level validation summary (%). A/T/R counts accepted, tied, and rejected harness proposals.

Dataset	RL update	Incumbent	Candidate	Change	Decision
OfficeQA	5	41.51	45.28	$+3.77$	Accept
OfficeQA	30	60.37	49.06	$-11.31$	Reject
AutomationBench	5	53.81	61.07	$+7.26$	Accept
AutomationBench	15	70.93	64.50	$-6.43$	Reject

Table 4: Same-checkpoint examples from the phase-level trajectories: the first accepted and first rejected proposal in each dataset. Scores are percentages; changes are percentage points.

Appendix figures & tables1 asset

Supplementary material from the paper’s appendix.

Appendix

Method	RL cap	H cap	Harness gate
Static agent	0	0	—
Harness-only	0	$B_{H}$	Yes
RL-only	$B_{W}$	0	—
SIA-style	$B_{W}$	$B_{H}$	Yes
WHALE-style	$B_{W}$	$B_{H}$	No
VACE	$B_{W}$	$B_{H}$	Yes

Appendix

Table 5: Optimization settings.

B_{W}

and

B_{H}

denote the planned RL and harness-proposal caps.

Explore similar work

WHALE: A Simple Recipe for Joint Harness-Weight Optimization
Aug 31, 2026Haechan Kim, Yoonho Lee, Gisang Lee +2Agent Harness Agentic Optimization
KRAFTON · KAIST · Stanford University
EnvACE: Internalizing Environment Dynamics via World Rehearsal for Agentic Reinforcement Learning
Aug 6, 2026Zishan Xu, Zhiyuan Yao, Yuxin Chen +9Agentic Reinforcement Learning Large Language Model Agents
Shanghai Jiao Tong University · Zhejiang University · National University of Singapore +4
PACE: Two-Timescale Self-Evolution for Small Language Model Agents
May 21, 2026Chen Ling, Pei Chen, Albert Guan +4Self-Evolution Large Language Model Agents
Amazon

Aug 31, 2026, cs.LGJ/K move · Enter open · S save

WHALE: A Simple Recipe for Joint Harness-Weight Optimization

Haechan Kim, Yoonho Lee, Gisang Lee, Chelsea Finn+1

KRAFTON · KAIST · Stanford University

Agent Harness Agentic Optimization Batch Recipe

Open paper AskPDF arXiv

Agent performance depends jointly on the model parameters and the executable harness code that manages context and control flow. Optimizing either component in isolation can leave the system bottlenecked by its frozen counterpart: weight updates can change which harness is effective, while harness updates can change which model capabilities are exposed. Existing joint-adaptation methods optimize weights and textual prompts but leave the broader harness fixed. We propose Weight-Harness Alternating LEarning (WHALE), a simple recipe that alternates two phases: updating the model under the current harness, then searching for a better harness under the updated model. We instantiate these two phases with online rejection-sampling fine-tuning and Meta-Harness, respectively. When to switch is a key design choice: to separate real improvements from noise without over-optimizing against a changing counterpart, WHALE uses either fixed phase durations or an adaptive patience rule over training signals. Using Qwen3.5-2B/4B agents across three domains (search question answering, mathematical reasoning, and chess puzzles), WHALE outperforms weight-only, harness-only, and Fast-Slow Training by 4.15-24.38 percentage points in best mean@8 accuracy. Either component can be the bottleneck: harness search matches peak weight-only accuracy with far fewer rollouts in SearchQA, but improves math accuracy only after a weight update. Small interleaved updates also outperform stagewise weight-then-harness optimization in accuracy and rollout cost. The code is available at https://github.com/krafton-ai/WHALE.

Abstract

Explore similar work

WHALE: A Simple Recipe for Joint Harness-Weight Optimization

EnvACE: Internalizing Environment Dynamics via World Rehearsal for Agentic Reinforcement Learning

PACE: Two-Timescale Self-Evolution for Small Language Model Agents