Latent Harness: Compiling Agent Behaviors into Post-Block Activation Shifts

Ran Li; Shimin Di; Zeyu Ren; Ling Yue; Jiachuan Wang; Lei Chen

cs.CLNov 1, 2026

Latent Harness: Compiling Agent Behaviors into Post-Block Activation Shifts

Authors: Ran Li, Shimin Di, Zeyu Ren, Ling Yue, Jiachuan Wang, Lei Chen

Organizations: Hong Kong University of Science and Technology · Southeast University · Rensselaer Polytechnic Institute · University of Tsukuba ·

Abstract

Harness engineering and self-evolving agents improve language-model agents quickly, yet their gains live outside the model. The harness is re-sent and re-paid on every call, and the standard route to internalization is a weight update that is costly to iterate and fixed once merged. We propose compiling the harness instead. Latent Harness distills an external condition into a rank-8 activation shift. The shift is applied at the output of every decoder block of a frozen backbone, at 0.03-0.05% of parameters. The condition is either an explicit process-level behavior instruction or the behavioral contract of an agentic harness. At inference the compiled bundle carries no harness text and adds no tokens. It exposes three runtime controls for injection strength, injection phase, and shift-space mixing. Compiled behaviors recover the source-harness gain in 8 of 10 preregistered settings. Where the explicit source pays the heaviest per-call token tax, the compiled form exceeds it. The same bundles transfer zero-shot across tasks (5 of 7 pairs positive) and compose in shift space, while stacked text instructions interfere destructively. Used as a search space, they lift budget-matched harness search by 1.45 times. On three agentic benchmarks the operator aligns policy to the harness con- tract. Under matched data, capacity, and effective strength it outperforms same-capacity LoRA, which needs 11.1 times the parameters to reach the same band, and it regresses nothing where the base policy is already strong. Across 1B-to-8B models the method skeleton is invariant while gain locations move with the model-task pair, which delimits where compiled harnesses should be deployed.

Figures & tables

Figure 1: Latent Harness compiles external conditions into post-block activation shifts. Teacher traces or strictly filtered success trajectories train a rank-8 shift at every block output of a frozen backbone. The bundle injects at run time with strength, phase, and mixing controls.

			ToolHop		SearchQA			API-Bank
Model	Method	Params (%)	Ans.	Path	Hotpot	2Wiki	Overall	Succ.	API.
Qwen3-4B	ADAS	0	40.00	44.99	28.67	28.00	28.33	57.31	60.99
	AgentSquare	0	29.23	39.46	29.33	30.00	29.67	47.82	46.10
	AFlow	0	31.28	43.81	30.67	29.33	30.00	45.72	41.13
	MaAS	0	42.05	51.80	30.67	30.00	30.33	62.74	63.83
	MermaidFlow	0	44.62	58.74	31.33	28.00	29.67	64.33	60.28

Table 1: Main results on the HarnessForge benchmark suite, following its baseline suite. Search-style rows search over agent harnesses, training-style rows update policy weights, and HarnessForge is the full co-evolution framework. Ours is evaluated under our controlled single-round alignment protocol, which uses the official round-3 harness pairing and evaluation pools disjoint from training, and which is protocol-distinct from the co-evolution rows above. Params (%) is the trainable fraction of backbone parameters (LoRA for training-style rows and HarnessForge, while search-style rows update no weights). Best in bold , second best underlined , per backbone.

Panel of Figure 2 — Figure 2: Policy compilation, efficiency, stability, and safety. (a) Capacity alignment on ToolHop. (b) Run-level repeatability. (c) Zero regression on API-Bank at 4B and 8B.

Setting	EXPL.	SOFT-PROMPT	TASK-ADAPTER	OURS	$\Delta_{L}$ ( $R$ )
M1 $\cdot$ ARC $\times$ COMPUTE	0.300	0.347	0.533	0.322	$+0.082$ (1.23)
M1 $\cdot$ ARC $\times$ CONCISE	0.380	0.376	0.533	0.362	$+0.122$ (0.83)
M1 $\cdot$ MBPP $\times$ TESTCONSULT	0.066	0.079	0.308	0.323	$+0.317$ (5.86)
M2 $\cdot$ MBPP $\times$ DECOMPOSE	0.410	0.508	0.562	0.661	$+0.619$ (1.66)
M2 $\cdot$ MBPP $\times$ COMPUTE	0.470	0.521	0.554	0.660	$+0.618$ (1.43)
M2 $\cdot$ MBPP $\times$ CONCISE	0.396	0.531	0.570	0.687	$+0.645$ (1.80)

Table 2: Controlled test-set comparison, one touch and preregistered. SOFT-PROMPT is parameter-matched and trained on the same distillation signal. TASK-ADAPTER is trained on gold labels at matched capacity.

\Delta_{L}

is ours over base, and

R=\Delta_{L}/\Delta_{E}

uses the explicit gains in Fig. 7 . All

\Delta_{L}

CIs exclude zero, and 8 of 10 enrolled settings meet the criterion

R\geq 0.5

Panel of Figure 3 — Figure 3: Instruction compilation. (a) Behavior recovery on the test set. (b) Token cost against accuracy.

Panel of Figure 4 — Figure 4: Three-chain attribution. (a) Accuracy against behavior markers (M2 $\times$ MBPP average). (b) Score recovery against marker recovery. (c) Harmful signatures compile symmetrically and transfer out of distribution.

Panel of Figure 5 — Figure 5: Reuse of compiled behaviors. (a) Cross-task transfer of frozen bundles. (b) Text stacking against shift-space composition. (c) Search gain at matched evaluation budgets.

Panel of Figure 6 — Figure 6: Mechanism. (a) Phase decomposition of the gain. (b) Strength response with plateau and cliff. (c) ToolHop with the harness removed, by configuration.

Appendix figures & tables11 assets

Supplementary material from the paper’s appendix.

Appendix

Code	Model	Layers $\times$ $d$	Trainable	Share
M1	Llama-3.2-1B-Instruct	16 $\times$ 2048	524,288	0.0424%
M2	Qwen3-4B	36 $\times$ 2560	1,474,560	0.0367%
M3	Llama-3.1-8B-Instruct	32 $\times$ 4096	2,097,152	0.0312%
v2	Llama-3.2-1B (base)	16 $\times$ 2048	524,288	0.0424%
–	Qwen3-8B (policy)	36 $\times$ 4096	2,359,296	0.0288%
Instruction tasks are BoolQ, ARC-Challenge, GSM8K, and MBPP.

Appendix

Table 3: Models and coverage. Rank-8 post-block adapters cover all layers and all tokens, with a SiLU bottleneck for instruction compilation and a linear map for policy compilation. The backbone is frozen, with every parameter verified identical to the original.

Figure 7: Screening forest (left) and the MBPP format-floor unlock (right).

Figure 8: MBPP marker displacements under explicit harnesses (left) and absolute marker transitions from base to the compiled form (right).

Figure 9: (a) Preregistered marker recoveries across settings. (b) Independent replication on the base 1B model.

Figure 10: (a) Harmful-signature transfer to SVAMP. (b) Transfer contrast between compiled behaviors and gold-label task adapters.

Figure 11: (a) The 25-point shift-space composition grid. (b) The three budget-matched search configurations.

Figure 12: (a) All ToolHop configurations, with runs as points. (b) Action counts per configuration. (c) Reproducibility self-check.

Figure 13: (a) Failure staging at high strength. (b) The same failure family at 4B and 8B.

Dimension	4B	8B	Verdict
ToolHop both gain	[email protected] reproduces 55/95 ( $+4.3$ )	multi-band: [email protected] $+4.6$ pt, [email protected] $+4.0$ pt	gain location moves with scale
SearchQA gain	none (Hotpot subset)	Overall $+7.5$ – $9.2$ pt (Hotpot $+11$ – $15$ )	direction flips with scale
Phase conclusion	both carries the stable gain ( $\approx$ 60% prompt-side)	gen $\approx$ both on ToolHop and both wins on SearchQA	task $\times$ scale dependent
LoRA comparison	same-capacity LoRA $\leq$ base everywhere	LoRA r11 (3.09 $\times$ params) matches Ours	Ours matches LoRA with $\leq$ 1/3 params
Scale safe interval	plateau 0.35–0.70 and cliff $\geq 0.8$	narrows to 0.25–0.5 and collapse $\geq 0.875$	same failure family, earlier onset
API-Bank	zero regression (bit-exact tie)	all configurations at or above base ([email protected] $+4.4$ pt)	zero regression across scales

Appendix

Table 4: Cross-scale account for Qwen3-4B against Qwen3-8B. The method skeleton is invariant. Gain locations (task

\times

phase

\times

scale) move with the model

\times

task pair.

Figure 14: (a) Gain locations across benchmarks and scales. (b) 8B ToolHop across strengths.

Figure 15: (a) 8B SearchQA by subset. (b) 8B API-Bank by configuration.

LLM Agent Harnesses

LLM agent harnesses are the software and runtime systems that organize an agent’s tools, context, state, execution, and safeguards.

See all 406 papers

Meta-Harness: End-to-End Optimization of Model Harnesses
Mar 30, 2026Yoonho Lee, Roshen Nair, Qizheng Zhang +3LLM Agent Harnesses Large Language Model-Guided Optimization
Stanford · KRAFTON · MIT
MESH-Harness: Self-Improving Agent Harnesses via Bandit-Guided Compositional Evolution
Oct 4, 2026Zhiwei Shang, Yu Huo, Mingrong Gong +6Agent Harness Optimization LLM Agent Workflow Optimization
The Chinese University of Hong Kong, Shenzhen · DeepWisdom · The Chinese University of Hong Kong +3
HarnessDev: Can LLMs Create and Evolve Their Own Agent Harness?
Sep 1, 2026Yuhao Wu, Jingyuan Zhang, Jiajun Shi +16LLM Agent Evaluation Agent Harness Optimization

Mar 30, 2026J/K move · Enter open · S save

Meta-Harness: End-to-End Optimization of Model Harnesses

Yoonho Lee, Roshen Nair, Qizheng Zhang, Kangwook Lee+2

Stanford · KRAFTON · MIT

LLM Agent Harnesses Large Language Model-Guided Optimization

Open paper AskPDF arXiv

The performance of large language model (LLM) systems depends not only on model weights, but also on their harness: the code that determines what information to store, retrieve, and present to the model. Yet harnesses are still designed largely by hand, and existing text optimizers are poorly matched to this setting because they compress feedback too aggressively. We introduce Meta-Harness, an outer-loop system that searches over harness code for LLM applications. It uses an agentic proposer that accesses the source code, scores, and execution traces of all prior candidates through a filesystem. On online text classification, Meta-Harness improves over a state-of-the-art context management system by 7.7 points while using 4x fewer context tokens. On retrieval-augmented math reasoning, a single discovered harness improves accuracy on 200 IMO-level problems by 4.7 points on average across five held-out models. On agentic coding, discovered harnesses surpass the best hand-engineered baselines on TerminalBench-2. Together, these results show that richer access to prior experience can enable automated harness engineering.

Abstract

LLM Agent Harnesses

Explore similar work

Meta-Harness: End-to-End Optimization of Model Harnesses

MESH-Harness: Self-Improving Agent Harnesses via Bandit-Guided Compositional Evolution

HarnessDev: Can LLMs Create and Evolve Their Own Agent Harness?