Necessity-Action Mismatch

• Necessity–action mismatch is a concept describing the gap between actions required by external constraints and those actually executed by agents such as AI systems, robots, and LLMs.
• It is quantified using metrics like neurosis scores, mismatch rates, and probe analyses across diverse domains including embodied AI, distributed systems, and tool-use in language models.
• This topic has significant implications for enhancing model-groundedness, ensuring safety, and improving coordination in multi-agent and multimodal systems.

Necessity-action mismatch denotes any systematic gap between the actions an agent—ranging from physical robots to LLMs—needs to take (necessitated by external demands, safety constraints, or task specifications) and the actions it actually executes. This phenomenon arises across disciplines: metaphysical analyses of physical law, epistemic requirements in distributed protocols, behavioral failures in embodied AI, and “knowing-doing” gaps in machine learning systems. The necessity–action mismatch has emerged as a central diagnostic for model-groundedness, safety, and correctness.

1. Formal Definitions Across Domains

Each research community frames necessity–action mismatch according to its model of agency and action:

• Embodied AI: Given system state $S_t$ and action set $A(S_t)$, an “external observer” prescribes a policy minimizing cost $c^*(S_t) = \min_{a\in A(S_t)} c(S_t, a)$; the agent’s chosen action $a_t$ generally deviates. The instantaneous mismatch is $M_t = c(S_t, a_t) - c^*(S_t) \ge 0$. Aggregating $M_t$ quantifies neurosis and accumulated inefficiency or safety violation (Howard, 12 Oct 2025).
• Distributed/Knowledge Systems: If $\varphi$ is a necessary condition for performing action $\alpha$, the Knowledge of Preconditions (KoP) principle requires that agent $i$ must know $\varphi$ whenever $A(S_t)$0 performs $A(S_t)$1. Any failure to ensure $A(S_t)$2 before action invokes a necessity–action mismatch (Moses, 2016).
• LLMs as Tool-using Agents: Let $A(S_t)$3 label the necessity of tool use by model $A(S_t)$4 on input $A(S_t)$5, and $A(S_t)$6 mark whether the tool is actually invoked. The mismatch rate $A(S_t)$7 quantifies the decoupling (Cheng et al., 13 May 2026).
• Omnimodal LLMs and Grounded Perception: Internal states may linearly encode that a sensory–textual conflict exists (necessity to reject), but outputs may still accept false premises—capturing representation–action gaps (Quang et al., 13 May 2026).
• Physics and Law Hierarchy: The Principle of Least Action (PLA) intermediates between metaphysical necessity (what must exist in all possible worlds) and physically enacted laws (actual motion), highlighting tension between modal necessities and realized system behavior (Terekhovich, 2015).

2. Theoretical Foundations and Principles

Necessity–action mismatch structurally arises whenever an agent’s “knowledge,” “internal representation,” or “grounding” is disjoint from its behavioral execution:

• Knowledge of Preconditions Principle (KoP): For any “conscious” action $A(S_t)$8 of agent $A(S_t)$9 in system $c^*(S_t) = \min_{a\in A(S_t)} c(S_t, a)$0, if state $c^*(S_t) = \min_{a\in A(S_t)} c(S_t, a)$1 is necessary for $c^*(S_t) = \min_{a\in A(S_t)} c(S_t, a)$2, then $c^*(S_t) = \min_{a\in A(S_t)} c(S_t, a)$3 (knowing $c^*(S_t) = \min_{a\in A(S_t)} c(S_t, a)$4) is also necessary for $c^*(S_t) = \min_{a\in A(S_t)} c(S_t, a)$5. This is formalized as:

$c^*(S_t) = \min_{a\in A(S_t)} c(S_t, a)$6

The KoP principle recasts task specifications involving world-states into explicit epistemic requirements (Moses, 2016).

• Modal Metaphysics of Physical Laws: The PLA exemplifies a mismatch in the categorization of necessity: metaphysical (all logics/worlds), general physical (structure of physics), and contingent physical (actual trajectories). The mismatch lies in the gap between the system’s modal requirements and empirical outcomes (Terekhovich, 2015).
• Internal Representation vs. Output Generation: In LLMs and embodied models, linearly decodable signals about necessary actions (e.g., “need a tool,” “premise is false”) may exist internally but are not propagated through the readout to ultimate behavior, producing knowing–doing or representation–action gaps (Cheng et al., 13 May 2026, Quang et al., 13 May 2026).

3. Empirical Manifestations and Metrics

Necessity–action mismatches manifest as observable errors or pathologies in agent behavior:

• Embodied Agents (AI Neurosis): Behaviors such as metric mismatch, tie-break thrashing, corridor thrashing, and policy oscillation all represent gaps between what actions are mandated by cost/law and what is performed. Instrumentation via neurosis score, $c^*(S_t) = \min_{a\in A(S_t)} c(S_t, a)$7, and related statistics surfaces both local and global mismatches (Howard, 12 Oct 2025).
• Tool-Use in LLMs: Mismatch rates of 26.5–54.0% (arithmetic) and 30.8–41.8% (factual QA) are observed, indicating a substantial fraction where an LLM “should” call a calculator or external API but does not, or vice versa. Analysis via linear probing reveals dissociation between the cognitive necessity signal and the execution signal at the final generation layer, with probe directions becoming orthogonal (Cheng et al., 13 May 2026).
• Multimodal LLMs: On benchmarks such as IMAVB, the necessity–action gap is evidenced by high probe accuracy for detecting misleading premises in hidden states (up to 86%) but much lower behavioral rejection rates (≤16.2% for vision, ≤6.6% for audio). The action gap remains modality-asymmetric and resistant to prompt engineering (Quang et al., 13 May 2026).

4. Mechanisms, Modalities, and Detectors

Necessity–action mismatches are characterized by domain-specific patterns and diagnostic tools:

• Modalities in Agents: Persistent metric–action mismatches (e.g., due to unmodeled slip), recurrent tie-break thrashing (route switching), and policy oscillation are classic patterns, each quantifiable with lightweight online detectors operating through windowed statistics on cost, flip rate, and plan consistency (Howard, 12 Oct 2025).
• Knowledge Protocols: In distributed systems, failure to ensure necessary epistemic preconditions (common or nested knowledge) before joint or sequential action guarantees a mismatch, as in Firing Squad and ATM scenarios (Moses, 2016).
• Probe-based Analysis: In LLMs, linear classifiers on hidden states (“probes”) can predict necessity labels, while measurement of cosine similarity between probes for cognition and execution quantifies their misalignment (Cheng et al., 13 May 2026).

5. Interventions and Mitigation Strategies

Research identifies both architectural and algorithmic strategies to reduce necessity–action mismatch:

• Escape Policies in Robotics: Commitment windows, margin-to-switch gating, temporal smoothing, canonical tie-breaks, and confidence-weighted arbitration are formalized heuristic interventions ensuring temporary congruence between plan and action, with overrides for safety or major gains (Howard, 12 Oct 2025).
• Destructive Testing: Genetic programming (GP) is leveraged to evolve adversarial world instances maximizing necessity–action mismatch, measured via custom neurosis and law-pressure scores; this discovers global failure modes that evade local heuristics (Howard, 12 Oct 2025).
• Representation–Action Coupling in LLMs: Recent work demonstrates that adding probe-guided logit adjustments can bridge the action gap, consistently increasing rejection rates on adversarial tasks without large loss to standard accuracy. Joint objectives, logit biasing, head fine-tuning, and representation-adapter modules are proposed for more robust translation of necessity signals to action (Quang et al., 13 May 2026, Cheng et al., 13 May 2026).
• Epistemic Protocol Design: In distributed protocols, explicit incorporation of knowledge requirements (KoP and its variants) ensures action only upon satisfaction of necessary conditions as epistemically ascertained, eliminating specification–execution discrepancies (Moses, 2016).

6. Broader Implications and Schematic Ontology

Necessity–action mismatch serves as an organizing framework across several disciplines:

• In physics, necessity-action relations underlie the hierarchy of laws—metaphysical, general physical, and system-specific—and their grounding in ontology and modal theory (Terekhovich, 2015).
• In AI systems, necessity–action mismatch diagnostics expose model brittleness, surface hidden failure regimes, and clarify the essential coupling required between perception/knowledge and operational output (Howard, 12 Oct 2025, Quang et al., 13 May 2026, Cheng et al., 13 May 2026).
• In multi-agent systems, rigorous epistemic analysis (KoP, common/nested knowledge) transforms vague “must-hold” requirements into concrete knowledge specifications, enhancing correctness and coordination (Moses, 2016).

A unifying schematic can be presented as a hierarchy:

Necessity Tier	Law/Mechanism	Agent Domain
Metaphysical Necessity	Combination/striving laws	Modal metaphysics, PLA (Terekhovich, 2015)
General Physical	Principle of Least Action	Physics, systemic law selection
Limited Physical	Newton, Maxwell, local rules	Mechanics, agent-specific behaviors
Epistemic Enforcement	KoP, C_G, nested knowledge	Multi-agent coordination (Moses, 2016)
Policy/Behavioral	Commitment, arbitration	Robotic/AI action selection (Howard, 12 Oct 2025)
Representation-Output	Probe-guided, logit bias	LLMs, tool use, multimodal alignment (Quang et al., 13 May 2026, Cheng et al., 13 May 2026)

Downward arrows link higher necessity to operational realization, with necessity–action mismatch emerging wherever transmission is impaired.

7. Open Problems and Future Directions

The persistence of necessity–action mismatch points to several enduring research challenges:

• Closing the Gap in LLMs: Mechanisms that consistently propagate internal necessity signals to outputs, especially across modalities, remain an open area—requiring advances in joint training regimes, output-head alignment, and possibly “meta-controller” architectures.
• Global versus Local Diagnoses: Local escape strategies can suppress symptoms of mismatch, but adversarially constructed worlds and failure modes (via GP or counterfactual analysis) remain indispensable for surfacing structural issues and guiding architectural overhaul (Howard, 12 Oct 2025).
• Integration of Epistemic Principles: Translating knowledge-theoretic principles into practical system engineering—ensuring every precondition-induced action actually aligns with system knowledge—demands further development, especially under uncertainty, partial observability, and real-time constraints (Moses, 2016).
• Modal Categorization of Physical Law: Clarifying the tiered necessity structure in physical and engineered systems offers deeper theoretical insight, and potentially new pathways for AI groundedness, safety, and coordination (Terekhovich, 2015).

Necessity–action mismatch thus remains a foundational yet broadly applicable analytic concept, structuring troubleshooting, theory, and design across physical law, distributed systems, and AI agent architectures.