Interaction Locality in Hierarchical Recursive Reasoning
Yosuke Miyanishi, Tetsuro Morimura
Abstract
Spatial reasoning requires both location-bound computation and location-invariant structure: agents must make local moves while preserving route, object, or constraint-level plans. We propose interaction locality, a task-geometry-aware framework for measuring whether information flow stays within nearby cells or semantic segments, or crosses them. We instantiate the framework with sparse-autoencoder feature ablations and finite-noise activation patching, with structural Jacobian and attention checks reported in the appendix, and apply it to HRM and TRM, two compact hierarchical and recursive reasoning models, on Maze-Hard, Sudoku Extreme, and ARC-AGI. Across these models, activation patching gives the clearest architectural fingerprint: high-level recurrent states tend to write information within nearby cells or same-segment units, while repeated recursive updates accumulate these local writes into broader solution structure. This pattern holds across maze paths, Sudoku constraints, and ARC-AGI object neighborhoods, with the strongest concentration in TRM. To test whether interaction locality extends beyond toy-yet-challenging grid benchmarks, we also apply it to MTU3D, a large-scale embodied 3D scene-grounding model. In this MTU3D setting, causal spatial locality appears primarily at the transition where visual scene features are handed to the downstream grounding module, rather than uniformly throughout the visual encoder. This contrast suggests that the local-to-global handoff observed in HRM and TRM is tied to explicit recursive reasoning dynamics, while embodied 3D models may concentrate causal spatial structure at module boundaries. Interaction locality turns the intuitive local-execution/global-planning story into a reproducible measurement framework for recursive and embodied spatial reasoning.
AI Impact Assessments
(1 model)Scientific Impact Assessment: "Interaction Locality in Hierarchical Recursive Reasoning"
1. Core Contribution
The paper proposes interaction locality, a measurement framework for assessing whether internal information flow in spatial reasoning models respects task-defined geometric neighborhoods (e.g., maze corridors, Sudoku boxes, ARC-AGI object regions). The framework is instantiated with three probe types: sparse autoencoder (SAE) feature ablations, finite-noise activation patching (the primary causal tool), and structural Jacobian/attention checks (appendix). It is applied to HRM and TRM—two compact hierarchical/recursive reasoning architectures—across Maze-Hard, Sudoku Extreme, and ARC-AGI, plus MTU3D, a 3D embodied grounding model on ScanNet.
The central finding is that H-level (high-level) recurrent states tend to write information *locally* within task-defined neighborhoods, while cross-cycle propagation accumulates these local writes into broader structure. This revises the naive "H = global, L = local" narrative. The MTU3D extension shows that causal spatial locality concentrates at the visual-to-grounding handoff rather than uniformly through the encoder, suggesting the local-to-global handoff is specifically tied to recursive reasoning dynamics.
2. Methodological Rigor
Strengths in methodology:
Weaknesses:
3. Potential Impact
The framework addresses a genuine gap: mechanistic interpretability for spatial reasoning models lacks a unified coordinate system for comparing locality across tasks and architectures. The idea of measuring information flow against task geometry is conceptually clean and could become a useful diagnostic tool.
Practical applications include: (1) diagnosing whether recursive reasoning models develop the intended local-to-global computation, (2) guiding locality-aware training objectives (though none are evaluated here), and (3) providing a common language for comparing spatial reasoning across domains (grids, 3D scenes, etc.).
However, impact is limited by: the niche scope of the models studied (HRM/TRM are compact research models, not widely deployed systems), the absence of training-time experiments showing that locality diagnostics actually improve model design, and the relatively descriptive nature of findings—the framework reveals properties but doesn't yet prescribe improvements.
4. Timeliness & Relevance
The paper is timely in several respects. Mechanistic interpretability is a rapidly growing field, and extending it beyond language models to spatial reasoning is valuable. The studied architectures (HRM, TRM) are very recent (2025), and the integration with embodied 3D models (MTU3D) addresses the growing interest in grounded spatial reasoning. The connection between recursive computation and spatial locality is relevant to ongoing debates about how compact models solve complex reasoning tasks.
However, the benchmarks remain "toy-yet-challenging" (the paper's own characterization), and the MTU3D extension, while promising, is limited to 30 scenes with primarily negative findings (no causal locality inside the encoder).
5. Strengths & Limitations
Key Strengths:
Key Limitations:
Overall Assessment
This is a carefully executed interpretability study that introduces a useful conceptual framework (interaction locality) and provides extensive empirical evidence across multiple tasks and architectures. The core insight—that what appears "global" in architecture labels may be causally local in information flow—is valuable. However, the impact is constrained by the narrow scope of studied models, modest sample sizes, purely diagnostic (rather than prescriptive) nature of findings, and the distance between the current grid-world applications and real-world deployment scenarios. The MTU3D extension is the most forward-looking contribution but yields primarily negative results within the encoder.
Generated May 21, 2026
Comparison History (25)
Paper 2 is likely higher impact: it introduces a broadly applicable measurement framework (interaction locality) with multiple instantiations (SAE ablations, activation patching, Jacobian/attention checks) and validates it across diverse models and domains, including ARC-AGI and a large-scale embodied 3D model (MTU3D), increasing breadth and real-world relevance. Its focus on mechanistic, reproducible characterization is methodologically rigorous and timely for interpretability of reasoning systems. Paper 1 shows strong performance gains and useful diagnostics, but is narrower (primarily an inference-time improvement technique for recursive models).
Paper 1 introduces a large-scale, practical benchmark evaluating LLMs on complex operations research tasks. Benchmarks bridging LLMs and scalable algorithm design have immense real-world applications in logistics, finance, and engineering, and typically attract high citation counts by driving sub-field progress. This will likely yield broader scientific and industrial impact compared to Paper 2's mechanistic interpretability framework, which, while rigorous, addresses a more niche aspect of spatial reasoning architectures.
Paper 1 addresses a critical and universal bottleneck in current AI research and deployment: debugging complex LLM agents at scale. By automating corpus-level trace diagnostics, it offers a highly practical, scalable tool that directly improves downstream agent performance (as evidenced by a 30.4pp gain). Its immediate real-world applicability and timeliness give it a broader and more immediate scientific impact compared to Paper 2, which focuses on a more niche, albeit rigorous, mechanistic interpretability framework for spatial reasoning models.
Paper 2 likely has higher scientific impact due to its methodological novelty and breadth: it proposes a general, reproducible measurement framework (interaction locality) for probing information flow in hierarchical/recursive reasoning and extends analysis across multiple benchmarks and an embodied 3D model, increasing cross-field relevance (interpretability, reasoning, robotics/vision). Its rigor (causal interventions via activation patching/feature ablations, auxiliary checks) supports strong, testable claims. Paper 1 is application-rich and timely, but is more of an engineering integration of existing agent/VLM components with limited generalization and modest success rates, likely narrowing impact.
MindLoom addresses a broadly impactful problem—synthesizing high-quality reasoning training data for LLMs—with a practical, open-sourced framework validated across nine benchmarks, multiple model families, and five STEM disciplines. Its compositional thought-mode engineering offers a novel, scalable methodology with immediate real-world applications in LLM training. Paper 2 introduces an interesting analytical framework (interaction locality) for understanding spatial reasoning in neural networks, but its scope is narrower, focused on interpretability of specific model architectures on grid-based tasks, with less immediate broad applicability and adoption potential.
Paper 1 introduces 'personality engineering,' a novel interdisciplinary methodology bridging AI and negotiation theory with broad applicability across social sciences, business, and AI agent design. It leverages the well-established interpersonal circumplex as a coordinate system, offering both theoretical contributions and practical implications for designing AI negotiation agents—a timely topic given LLM proliferation. Paper 2, while technically rigorous, addresses a narrower question about information flow locality in specialized reasoning models on grid-based tasks, with more limited audience and applicability beyond the interpretability/mechanistic analysis community.
Paper 2 introduces a novel interdisciplinary approach by integrating Prospect Theory into strategic classification, addressing a critical flaw in traditional ML models that assume perfect rationality. This bridging of machine learning and behavioral economics offers broader real-world applicability (e.g., in finance, hiring, public policy) compared to Paper 1, which focuses on a highly specific mechanistic interpretability framework for spatial reasoning models. Paper 2's potential to influence both AI fairness/robustness and behavioral science gives it a broader and higher estimated scientific impact.
Paper 2 likely has higher impact: it targets an urgent, widely relevant safety/alignment failure mode (sycophancy) in mainstream LLM deployments, proposes a practical mitigation using off-the-shelf persona vectors (no labeled sycophancy data), and reports strong performance with less accuracy tradeoff than CAA. Its finding that sycophancy is not a single steerable direction but relates to broader persona geometry could influence steering, evaluation, and interpretability work across many models and applications. Paper 1 is novel but narrower, focused on specialized spatial/recursive reasoning benchmarks and analysis tooling.
Paper 2 introduces a novel measurement framework ('interaction locality') with broader applicability across multiple reasoning domains (mazes, Sudoku, ARC-AGI, 3D embodied reasoning). It provides mechanistic interpretability insights into how hierarchical and recursive models process spatial information, bridging toy benchmarks and real-world embodied AI. Paper 1, while addressing a valid problem in RL for fighting games, presents incremental findings (learned timing matches fixed frame skips but doesn't ensure robustness) in a narrow application domain. Paper 2's framework has greater potential for cross-field impact in interpretability, spatial reasoning, and architecture design.
Paper 2 has higher impact potential: it identifies a broadly relevant, practically urgent failure mode in self-improving LLM agent systems (“library drift”), provides a clear causal trigger, introduces actionable diagnostics, and demonstrates a substantial, quantified fix with multiple ablations—strong methodological rigor and immediate real-world applicability. Its governance recipe and instrumentation can generalize across many agent/tooling frameworks, affecting reliability and long-horizon autonomy. Paper 1 offers a useful interpretability measurement framework, but it is narrower in application (spatial/recursive models) and its primary contribution is analytic rather than a widely deployable reliability improvement.