Measuring Harness-Induced Belief Divergence in Multi-Step LLM Agents

🕒 Published (v1): 2026-07-05 22:22 UTC · Source: Arxiv · link

Why this paper was selected

Directly measures how harness design (context/actions/repair) distorts agent beliefs; unique eval angle

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

Standard agent benchmarks report only terminal task success, treating the execution harness as neutral infrastructure. This paper shows that the harness systematically reshapes an agent's intermediate multi-step beliefs — about risk, recoverability, and failure mode — even when the task, environment, and base LLM are identical. It introduces a formal belief-rollout diagnostic, a decomposed divergence measure, and a no-training alignment protocol (BIWM) to make these harness-induced belief shifts auditable.

Problem

Agent evaluation collapses multi-step reasoning into a single outcome label, masking the fact that harness components — observation filtering, action gating, repair policies, verification masks, and logging — alter the evidence stream available to the LLM at each step. Changing the harness can leave final success unchanged while significantly distorting the intermediate beliefs (risk state, failure-mode expectation, recoverability estimate) that govern subsequent decisions. No prior work instruments this effect directly or provides a controlled measurement protocol for it.

Method

Belief-rollout diagnostic. For each task, harness \(H\), and horizon \(K \in \{1,3,5,8\}\), the LLM is elicited to produce a structured JSON belief state \(b \in \mathcal{B}\) at each rollout step. The belief space \(\mathcal{B}\) encodes: ordinal categorical fields (task progress \(p \in \{1\ldots5\}\), risk state \(r\), recoverability \(c\)), a discrete failure-mode label \(f \in F\), constraint sets \((K, S, V)\), numeric forecasts \((u, q_{\text{succ}}, q_{\text{fail}}, e_{\text{repair}}, m_{\text{hor}}, \rho_{\text{risk}}, e_{\text{cost}})\), and a recommended next action \(a\). The \(K\)-step rollout process is first-order on the augmented state \((B_t^H, h_t)\): $\(B_{t+1}^H \sim P_M^B\!\bigl(\cdot \mid T, B_t^H, \Gamma_H(h_t, A_t^H)\bigr)\)$ where \(\Gamma_H\) packages the harness-canonicalized observation, verifier signal, gated action, repair history, and log.

Cross-harness divergence. The scalar \(D_{\text{belief}}(H_a, H_b; K)\) is a weighted sum of five component distances over the final belief states: categorical (\(D_{\text{cat}}\), normalized ordinal), failure-mode (\(D_{\text{fail}}\), 0/1 indicator), constraint-set (\(D_{\text{set}}\), Jaccard complement), numeric (\(D_{\text{num}}\), field-wise normalized MAE), and action (\(D_{\text{act}}\), first-8-token disagreement).

Arrival/growth decomposition. \(D_{\text{belief}}\) decomposes exactly as: $\(D_{\text{belief}}(H_a, H_b; K) = w_{\text{arr}}\,D_{\text{arrival}}(H_a, H_b; K) + w_{\text{gro}}\,D_{\text{growth}}(H_a, H_b; K)\)$ \(D_{\text{arrival}}\) captures immediate interface mismatch (constraint content, action format); \(D_{\text{growth}}\) captures horizon-dependent drift in risk, failure-mode, and numeric forecast fields.

Harness family. Six harnesses isolate distinct mechanisms: raw reference (\(H_0\)), structured (parsed tracebacks + explicit states), risk-gated (blocks high-risk actions, withholds outcomes), repair-heavy (compresses failure paths via retry/rollback), verification-selective (partial state checking), cost-aware (omits expensive checks under budget pressure).

BIWM protocol. Belief-Invariant World-Modeling is a no-training inference-time protocol that: (1) canonicalizes observations via a shared schema embedding \(\iota_O\); (2) logs censored branches with blocked-action metadata rather than silently omitting them; (3) expands repair traces to expose the suppressed failure path; (4) records verification masks (checked vs. unverified states); (5) executes risky branches in shadow before gating; (6) aligns belief trajectories across harness views.

Key Contributions

• Belief-rollout diagnostic: structured \(K\)-step elicitation protocol for controlled cross-harness comparison, holding task, LLM, schema, and decoding fixed
• \(D_{\text{arrival}}/D_{\text{growth}}\) decomposition: separates immediate interface effects from horizon-dependent belief drift; proves a constraint-set floor lemma (\(D_{\text{belief}} \geq w_{\text{set}}\) when constraint sets are disjoint) and a censoring-induced growth monotonicity theorem under three propagation conditions
• BIWM: a no-training protocol that makes harness censorship visible to the belief update rather than opaque, reducing belief divergence across harness views
• Empirical demonstration: shows that blocked actions, compressed repairs, selective verification, and cost-aware evidence pruning frequently preserve terminal success while altering intermediate beliefs that govern later decisions
• Methodological argument: establishes harness design as an experimental variable in agent evaluation, not an implementation detail

Results

• Between the raw reference harness \(H_0\) and risk-gated harness \(H_2\), \(D_{\text{growth}}\) grows from 0.012 at \(K=1\) to 0.037 at \(K=8\), demonstrating horizon-dependent belief accumulation
• BIWM reduces \(D_{\text{growth}}\) by approximately 1.65× relative to naive belief averaging across harness views (Figure 2B)
• Blocked actions, compressed repairs, selective verification, and cost-aware evidence pruning empirically "often preserve terminal success while changing the beliefs that drive later decisions" across controlled coding tasks and public-benchmark stress tests
• (Full quantitative tables not present in the provided text excerpt)

Limitations

• \(D_{\text{belief}}\) satisfies symmetry and boundedness but not the triangle inequality, making it a divergence rather than a metric; cross-harness comparisons are pairwise only
• Belief states are elicited via prompted JSON generation, not extracted from internal model representations; elicitation fidelity depends on prompt and schema design
• Theorem 1 (growth monotonicity) requires condition (C3) — non-negative expected growth increment at each step — which is asserted as a sufficient condition and may not hold universally
• Component weights \(w\) are fixed hyperparameters shared across all tasks, horizons, and harness comparisons; the paper does not analyze sensitivity to weight choices
• BIWM requires harness-level instrumentation (shadow execution, blocked-action logging, verification masks) that may not be available in black-box or production harnesses
• Evaluation is bounded to controlled coding tasks; generalization to web navigation, desktop control, or tool-use domains is asserted but not fully demonstrated in the provided text

Relevance to Agentic AI / LLM Agents

This paper directly addresses a blind spot in LLM agent evaluation: the assumption that the execution harness is neutral scaffolding rather than a variable that shapes agent cognition. For practitioners building or evaluating multi-step agents, it establishes that benchmark scores can be harness-confounded — the same agent achieving the same success rate may reason through fundamentally different belief trajectories depending on how the scaffold handles failures, verifications, and risky branches. The BIWM protocol is immediately applicable to any agentic system that uses structured JSON outputs, offering a no-training path to making intermediate world-model states auditable and comparable. The censorship finding — that safety guardrails blocking risky actions also censor information, potentially causing agents to infer that dangerous branches simply do not exist — is a significant safety concern for agentic systems with risk-gated scaffolds.