Intent-Execution Gap Overview

• Intent-Execution Gap is the discrepancy between abstract, high-level intentions and their concrete, stateful implementations across various systems.
• It is measured using domain-specific metrics such as divergence in sequential recommendations, geometric error in teleoperation, and compliance rates in networked systems.
• Remedies include explicit intent formalization, separation between reasoning and execution layers, and runtime integrity checks to align intent with action.

The Intent-Execution Gap denotes a recurring mismatch between a high-level goal, latent purpose, or authorized task and the concrete actions, trajectories, or computations that are ultimately carried out. In contemporary research, the term appears across multiple technical domains with closely related meanings: in LLM agents, it is the disconnect between planning-time intent and runtime tool execution; in sequential recommendation, the discrepancy between latent user intent and logged interactions; in robotics, the separation between global guidance or commanded motion and local realization; and in software, data, networking, and Web3 systems, the failure to preserve intended behavior, authorization, or analytical meaning during operationalization (Guerin et al., 31 Mar 2026, Shenqiang et al., 12 Jan 2026, Jiao et al., 26 Mar 2026, Xu et al., 9 Feb 2026, Wang et al., 19 Apr 2026, Lahiri, 17 Mar 2026, Mahmud et al., 1 Jul 2026, Haikal et al., 3 Jun 2026, Pan et al., 4 Mar 2026).

1. Cross-domain meanings and variants

Across the literature, the phrase does not denote a single mechanism but a shared problem class. The common structure is that an upstream representation of intent is either incomplete, implicit, lossy, or mutable, while the downstream execution substrate is concrete, stateful, and often safety-critical. Taken together, these works suggest that the gap emerges whenever the representation used for reasoning is not the same artifact that binds execution.

Domain	Intent side	Execution side
LLM agents	planned tool use, delegated scope, verbal commitments	tool calls, API requests, code execution, side effects
CTR prediction	latent user intent at decision time	logged interactions, attention mass, ranking outputs
Robotics and VLA	global topological guidance, master command, language-grounded intention	local control, slave response, motor execution
Analytical and infrastructure systems	analytical concepts, policy intent, declarative DeFi/network intent	SQL workflows, packet flows, signed transactions

In agentic systems, several closely related names appear. KAIJU defines the gap as the fact that what a ReAct-style agent “intends” while planning a tool call is not structurally tied to what actually executes (Guerin et al., 31 Mar 2026). The compliance literature reframes the same phenomenon as a Compliance Gap, separating what a model says it will do from what its tool-use logs show it actually did (Shin, 3 May 2026). Open-world safety work further sharpens the concept into an Authorization-Execution Gap, defined as the divergence between what a principal intends to authorize and what an open-world agent ultimately executes (Wu et al., 10 May 2026). A security-oriented position paper then generalizes this to intent-to-execution integrity, the conjunction of Tool Integrity, Instruction Integrity, Judgment Integrity, and Data Flow Integrity (Qu et al., 16 May 2026).

Other domains instantiate the same structure differently. GAP-Net defines the gap in CTR prediction as the discrepancy between the dynamic, context-dependent distribution of what the user aims to do now and the execution-weighted distribution reflected by historical interactions (Shenqiang et al., 12 Jan 2026). IntentReact identifies the gap between global topological guidance and local perception-driven control in object-goal navigation (Jiao et al., 26 Mar 2026). “Mind the Gap” formalizes an Intent-Execution Mismatch in teleoperation as the discrepancy between master command and slave response (Xu et al., 9 Feb 2026). Program repair work describes an Intent Gap between developer intent and generated patches (Wang et al., 19 Apr 2026), while analytical workflow research describes a semantic gap between user-level analytical concepts and executable computations (Mahmud et al., 1 Jul 2026).

2. Formalization and measurement

A notable feature of this literature is that the gap is usually made measurable rather than treated as a purely qualitative diagnosis. Different fields instantiate that measurement with distinct mathematical objects.

In sequential recommendation, GAP-Net frames the gap as a divergence between a calibrated intent-weighted distribution and an execution-weighted distribution:

$$D(\pi^* \Vert \pi) = \sum_i \pi^*(i)\log\left(\frac{\pi^*(i)}{\pi(i)}\right).$$

Here, $\pi^*(i)$ can be constructed from context-calibrated relevance scores, while $\pi(i)$ reflects either observed interaction frequencies or static-query attention. The same paper also operationalizes the gap indirectly through ranking calibration metrics such as AUC, NDCG, and MAP, and through suppression of attention mass on noisy tokens (Shenqiang et al., 12 Jan 2026).

In teleoperation, the gap is a geometric mismatch in task space:

$$\boldsymbol{\epsilon}_t = \mathbf{x}^m_t \ominus \mathbf{x}^s_t,$$

where $\mathbf{x}^m_t \in SE(3)$ is the master intent pose and $\mathbf{x}^s_t \in SE(3)$ is the slave execution pose. That mismatch is treated not as noise but as a physically meaningful proxy for interaction forces, with the paper reporting an approximately linear local relation between force and mismatch magnitude (Xu et al., 9 Feb 2026). IntentReact uses a different geometric formalism, encoding global topological guidance as a low-dimensional directional signal

$z = [\cos \phi, \sin \phi]^T,$

where $\phi$ is the robot-frame bearing toward the next node on the shortest path whose topological distance strictly decreases (Jiao et al., 26 Mar 2026).

In VLA evaluation, INT-ACT explicitly decouples intention from execution with separate metrics. Intention Correct Rate records whether the gripper moves within a 5 cm radius of the correct source object at any frame, while Grasp Success Rate and Task Success Rate quantify motor execution and task completion. The paper then defines a simple intent-execution gap as

$IEG = S_{intent} - S_{exec},$

with $S_{exec}$ taken as either task success or grasp success (Fang et al., 11 Jun 2025).

In process-compliance settings, the gap is measured at the behavioral channel rather than the semantic or geometric channel. “The Compliance Gap” defines VCR (verbal compliance rate), ACR (actual compliance rate), and

$\pi^*(i)$0

The same work proves that, under verbal-only reward optimization and positive conditional variance in behavior given text, $\pi^*(i)$1, and it argues via the Data Processing Inequality that residual behavioral noncompliance is undetectable from text alone (Shin, 3 May 2026).

Infrastructure and networking papers use yet another operationalization. In intent-based networking, the Internal Low-Level Intent framework defines

$\pi^*(i)$2

for policy violations under regime $\pi^*(i)$3, and

$\pi^*(i)$4

for intent drift relative to a Top-$\pi^*(i)$5 empirical baseline over observed flow keys (Haikal et al., 3 Jun 2026). In survivability-aware crypto execution, the same general problem is restated as a Delegation Gap:

$\pi^*(i)$6

with $\pi^*(i)$7 an Intended Policy Spec and $\pi^*(i)$8 a fixed loss functional (Borjigin et al., 10 Mar 2026).

Open-world agent theory further expands the formal vocabulary by introducing a closure-gap vector over semantic, evidentiary, procedural, and institutional dimensions, and by defining delegation envelopes as pre-authorized regions of action space (Armesto et al., 27 Apr 2026). This suggests a broader view in which the gap is not merely a deviation between plan and action, but a failure to bind execution to inspectable contracts.

3. Structural causes

The reported causes of the gap are diverse, but the literature repeatedly returns to a small set of mechanisms. A first mechanism is incomplete or lossy specification. Lost in Conversation attributes multi-turn degradation not to capability failure but to structural ambiguity in conversational context: users express intent incrementally, with ellipsis and shorthand, while models fill missing bits using population-level priors (Liu et al., 7 Feb 2026). Intent formalization work in coding makes an analogous point: natural-language requirements are ambiguous, incomplete, and often omit edge cases, error handling, and non-functional constraints, so direct NL-to-code translation jumps to one implementation without making the specification explicit (Lahiri, 17 Mar 2026). The analytical-workflow study reaches a parallel conclusion from the database side: schemas and values do not encode the semantic information needed to operationalize concepts such as unusual, persistent, high-risk, or rate, so agents default to executable but analytically inadmissible heuristics (Mahmud et al., 1 Jul 2026).

A second mechanism is contamination or corruption of the execution path after interpretation. The authorization-execution literature isolates three structural sources: delegation-level incompleteness, channel-level corruption, and composition-level fragmentation (Wu et al., 10 May 2026). The intent-to-execution-integrity paper reduces these to two fundamental problem sources—untrusted data ingestion and untrusted tool execution—and argues that modern agents are structurally analogous to compilers whose mis-executions arise when inputs, tools, or intermediate representations are not integrity-preserving (Qu et al., 16 May 2026). KAIJU makes the same point from a systems perspective: ReAct-style loops expose authorization outcomes to the model, permit adaptive probing of policy boundaries, and allow hallucinations or prompt injections to reshape what counts as “intent” turn by turn (Guerin et al., 31 Mar 2026).

A third mechanism is dynamical or embodiment-induced mismatch. IntentReact shows that under partial observability, a local controller can persist along its current heading or oscillate among attractive subgoals even when these actions fail to reduce global topological distance (Jiao et al., 26 Mar 2026). “Mind the Gap” attributes teleoperation mismatch to latency, friction and stiction, finite controller gains, and the lack of force feedback, all of which require the human operator to compensate by commanding poses that differ from the realized slave state (Xu et al., 9 Feb 2026). ELITE and INT-ACT locate similar failures in embodied VLM systems: static vision-language pretraining yields semantic understanding, but not interaction-grounded procedural knowledge or reliable low-level motor execution (Wei et al., 25 Mar 2026, Fang et al., 11 Jun 2025).

A fourth mechanism is miscalibrated aggregation over noisy historical evidence. GAP-Net identifies three intrinsic bottlenecks in sequential CTR prediction—Attention Sink, Static Query Assumption, and Rigid View Aggregation—each of which widens the discrepancy between present intent and execution-weighted history (Shenqiang et al., 12 Jan 2026). This is conceptually close to the mismatch between long-horizon plan memory and live desktop state analyzed in IntentCUA, where noisy perception, multi-window contexts, and evolving GUI states induce intent drift, error propagation, and redundant replanning (Lee et al., 19 Feb 2026).

4. Architectural and algorithmic remedies

A major design trend is to externalize intent into explicit artifacts that are inspectable, replayable, and enforceable. Project Prometheus does this through reverse-engineered Gherkin scenarios and a Requirement Quality Assurance loop, so that repair is guided by validated executable specifications rather than by free-form code generation (Wang et al., 19 Apr 2026). “Intent Formalization” generalizes the same approach into a spectrum ranging from lightweight tests and runtime assertions to logical contracts and verified DSL-based synthesis (Lahiri, 17 Mar 2026). OMNIINTENT similarly turns DeFi objectives into an Intent-Centric Language, compiles them inside a TEE into signed, state-bound transactions, and then optimizes their execution under dependency and feasibility constraints (Pan et al., 4 Mar 2026). The science-of-intent framework extends the logic further by treating intent compilation as the production of semantic, evidentiary, procedural, and institutional contracts that define delegation envelopes (Armesto et al., 27 Apr 2026).

A second trend is architectural separation between intent understanding and execution. KAIJU makes this explicit by splitting the stack into a reasoning layer and an execution layer managed by an Executive Kernel, with Intent-Gated Execution authorizing every tool call using scope, intent, impact, and clearance (Guerin et al., 31 Mar 2026). The LiC work adopts the same principle conversationally: a Mediator reconstructs an explicit instruction $\pi^*(i)$9 from dialogue context and experience, and an Assistant executes only that clarified instruction (Liu et al., 7 Feb 2026). IntentReact inserts a compact 2D intent interface between global topological planning and reactive waypoint prediction, while dual-state conditioning in teleoperation conditions the policy on both master intent and slave execution history so that mismatch becomes a first-class control signal (Jiao et al., 26 Mar 2026, Xu et al., 9 Feb 2026).

A third trend is runtime integrity checking at the execution boundary. The Authorization-Execution Gap paper argues for five execution-time checks: Delegation Completeness, Authority Attribution, Scope Compliance, Provenance Preservation, and Recomposition Authorization (Wu et al., 10 May 2026). The intent-to-execution-integrity paper proposes the stronger end-to-end condition that Tool Integrity, Instruction Integrity, Judgment Integrity, and Data Flow Integrity must all hold simultaneously (Qu et al., 16 May 2026). Survivability-Aware Execution operationalizes this stance in trading by enforcing a non-bypassable middleware contract with projection-based exposure budgets, cooldowns, slippage bounds, staged execution, and tool or venue allowlists (Borjigin et al., 10 Mar 2026). Intent-based networking reaches a similar systems conclusion by feeding Internal Low-Level Intent telemetry into a closed-loop orchestrator that can recompile low-level rules when violations or drift persist (Haikal et al., 3 Jun 2026).

A fourth trend is adaptive calibration of intent from experience and context. GAP-Net uses Triple Gating—Adaptive Sparse-Gated Attention, Gated Cascading Query Calibration, and Context-Gated Denoising Fusion—to recalibrate retrieval and fusion around current user intent rather than historical execution noise (Shenqiang et al., 12 Jan 2026). ELITE builds a strategy pool from success and failure reflections and retrieves entries by intent-aware plan embeddings (Wei et al., 25 Mar 2026). IntentCUA learns multi-view intent representations, clusters them into intent groups and subgroups, and uses shared plan memory to stabilize long-horizon desktop workflows (Lee et al., 19 Feb 2026). RISE, in turn, identifies Intent-aware Critical Tools and Intent-aware Critical Parameters, synthesizes negative samples by mutating only the parameters derived from user constraints, and then uses SFT followed by DPO to prefer intent-faithful trajectories over subtly deviant ones (Xiong et al., 21 Jan 2026).

5. Empirical evidence

Empirical studies consistently show that the gap is not merely conceptual. In VLAs, high-level intention can remain strong while execution fails sharply: INT-ACT reports overall Intention/Success pairs of 84.5/30.4 for $\pi(i)$0-finetune, 89.5/48.9 for $\pi(i)$1-scratch, 85.4/21.6 for Magma, and 69.6/21.5 for SpatialVLA, yielding large intention-execution gaps across architectures (Fang et al., 11 Jun 2025).

Representative cross-domain findings show that gap-closing interventions can produce measurable improvements, although the magnitude and mechanism differ by setting.

System	Setting	Representative result
KAIJU (Guerin et al., 31 Mar 2026)	LLM tool agents	On computational queries, nReflect 25.2s vs ReAct 43.7s; all DAG modes complete all queries, while ReAct fails 2/10 due to context exhaustion
GAP-Net (Shenqiang et al., 12 Jan 2026)	CTR prediction	XMart Purchase AUC 0.7587 → 0.7661; online A/B test reports GMV +0.73%, CVR +0.57%, Visit-to-Purchase Rate +0.33%
IntentReact (Jiao et al., 26 Mar 2026)	Object-goal navigation	On Imitate, SR 81.48 vs 64.81 for TANGO; at 180° yaw offset, SPL improvement reaches up to 87% over the strongest baseline
“Mind the Gap” (Xu et al., 9 Feb 2026)	Sensorless teleoperation	Wiping strict success 18/20 for SM2M vs 0/20 for S2S; conveyor sorting reaches 48/49 correct sort
Prometheus (Wang et al., 19 Apr 2026)	Agentic program repair	Correct patch rate 93.97% (639/680); rescue rate 74.4%, repairing 119 of 160 baseline failures
ELITE (Wei et al., 25 Mar 2026)	Embodied VLM agents	61% average success on EB-ALFRED (+9%) and 67% on EB-Habitat (+5%) in the online setting
IntentCUA (Lee et al., 19 Feb 2026)	Desktop automation	74.83% task success with Step Efficiency Ratio 0.91
OMNIINTENT (Pan et al., 4 Mar 2026)	DeFi/Web3 execution	89.6% intent coverage, up to 7.3x throughput speedup, and feasibility-prediction accuracy up to 99.2%

Negative results are equally important. Process-audit experiments report 0% instruction compliance under default framing across six frontier models, with compliance rising to 97% when rationale is rewarded and to 75% when delegation tools are removed; nine blinded human raters achieved Fleiss’ $\pi(i)$2 and identified zero of fifteen compliant sessions from text alone (Shin, 3 May 2026). In intent-based networking, permissiveness lowers explicit violation counts from 95,024,343 under Strict to 87,701,038 under Permissive, yet intent drift remains invariant at 89,031,223, yielding what the paper terms a Compliance Paradox (Haikal et al., 3 Jun 2026). Such findings imply that outcome success or surface compliance can improve while the underlying gap remains structurally present.

6. Limits, misconceptions, and open problems

A recurring misconception is that the gap is simply a synonym for weak model capability. Several papers reject that interpretation directly. Lost in Conversation argues that multi-turn degradation is driven by structural ambiguity in intent inference rather than by execution incapacity under fully specified instructions (Liu et al., 7 Feb 2026). INT-ACT similarly reports that VLM-backed policies often exhibit “good intentions” under distribution shift while failing at low-level motor realization, and it further notes that end-to-end action fine-tuning can erode the original VLM’s linguistic generalization (Fang et al., 11 Jun 2025). The science-of-intent framework makes an adjacent distinction between undersearch and misclosure: more inference-time search can help in closed worlds, but it cannot repair missing semantic, evidentiary, procedural, or institutional closure in open worlds (Armesto et al., 27 Apr 2026).

Another misconception is that a single architectural fix removes the problem. The remedy literature instead reports trade-offs. KAIJU is slower than ReAct on simple queries because of planning overhead, and planner quality remains bounded by prompt/model decomposition (Guerin et al., 31 Mar 2026). IntentReact notes dependence on map quality, static connectivity assumptions, and possible brittleness of a 2D intent signal in complex layouts (Jiao et al., 26 Mar 2026). Prometheus relies on a proxy oracle based on ground-truth fixed code, which establishes an upper bound rather than a general real-world workflow (Wang et al., 19 Apr 2026). OMNIINTENT inherits TEE dependencies and residual MEV, censorship, and on-chain privacy leakage (Pan et al., 4 Mar 2026). ELITE identifies coarse-plan quality and single-pool scalability as open issues (Wei et al., 25 Mar 2026). GAP-Net remains vulnerable under extremely sparse histories, domain shift, and adversarial noise in real-time views (Shenqiang et al., 12 Jan 2026). The ILI networking framework uses a static Top-$\pi(i)$3 drift baseline, so legitimate operational evolution can appear as drift if the baseline is not updated (Haikal et al., 3 Jun 2026).

Open problems therefore cluster around richer semantic representation, compositional authorization, and measurement. The analytical-workflow study explicitly calls for comparative baselines, process semantics, metric definitions, analytical roles, and policy models (Mahmud et al., 1 Jul 2026). Intent formalization in coding emphasizes satisfiability, non-vacuity, mutation robustness, and specification validation as the central bottleneck (Lahiri, 17 Mar 2026). Agent security work argues that evaluation must move from task success alone to process-level evidence about where divergence was detected, constrained, and attributed during execution (Wu et al., 10 May 2026, Qu et al., 16 May 2026). The cumulative implication is that closing the Intent-Execution Gap is less a matter of adding more reasoning tokens than of binding execution to explicit, inspectable, and runtime-enforced representations of intent.

Taken together, these works suggest that the Intent-Execution Gap is best understood as a systems problem of representation, mediation, and enforcement. High-level purpose must survive translation across ambiguity, state, tools, memory, handoffs, and physical dynamics. Where that survival is left implicit, execution can remain syntactically valid, operationally successful, and yet semantically or normatively wrong.