LACO: Adaptive Latent Communication for Collaborative Driving
Tianhao Chen, Yuheng Wu, Dongman Lee
Abstract
Collaborative driving aims to improve safety and efficiency by enabling connected vehicles to coordinate under partial observability. Recent approaches have evolved from sharing visual features for perception to exchanging language-based reasoning through foundation models for behavioral coordination. Though communicating in language provides intuitive information, it introduces two challenges: high latency caused by autoregressive decoding and information loss caused by compressing rich internal representations into discrete tokens. To address these challenges, we analyze latent communication in collaborative driving under inherent limitations of multi-agent settings. Our analysis reveals agent identity confusion, where direct fusion of latent states entangles decision representations across vehicles. Motivated by this, we propose LACO, a training-free \textbf{LA}tent \textbf{CO}mmunication paradigm that seamlessly adapts pretrained driving models to collaborative settings. LACO introduces Iterative Latent Deliberation (ILD) for latent reasoning, Cross-Horizon Saliency Attribution (CHSA) for communication-efficient information selection, and Structured Semantic Knowledge Distillation (SSKD) to stabilize ego-centric decision making. Closed-loop experiments in CARLA show that LACO notably reduces communication and inference latency while maintaining strong collaborative driving performance.
AI Impact Assessments
(1 model)Scientific Impact Assessment: LACO: Adaptive Latent Communication for Collaborative Driving
1. Core Contribution
LACO introduces a paradigm shift in collaborative driving communication by replacing language-based or visual-token-based inter-vehicle communication with selective KV cache exchange from transformer-based Vision-Language-Action (VLA) models. The paper identifies a critical phenomenon—agent identity confusion—where naive full-depth KV cache fusion causes the ego vehicle to over-attend to a collaborator's deep-layer representations, effectively hijacking its control policy. This is a genuine and well-characterized failure mode that had not been previously documented in the collaborative driving literature.
The framework comprises three components: (1) Iterative Latent Deliberation (ILD), which performs latent reasoning through iterative forward passes without autoregressive token generation; (2) Cross-Horizon Saliency Attribution (CHSA), which prunes spatially redundant tokens based on attention-derived saliency scores; and (3) Structured Semantic Knowledge Distillation (SSKD), which restricts communication to shallow transformer layers where representations remain globally informative but not yet entangled with ego-specific control synthesis.
The training-free nature is a significant practical advantage—it allows direct deployment on pretrained VLA models without fine-tuning, which is critical given the cost of training large driving models.
2. Methodological Rigor
The motivation study is the paper's strongest methodological contribution. The attention entropy analysis revealing a U-shaped trajectory across layers (global parsing → ego-centric contraction → control synthesis resurgence) provides principled justification for shallow-layer-only fusion. The quantitative characterization of spatial attention sparsity (~30% of tokens capturing most attention mass) similarly grounds the CHSA design.
However, several concerns arise:
The ablation studies are reasonably thorough, covering each component, distillation depth, retention rate, and latent step count. The finding that 10% depth and 30% retention rate are broadly optimal across architectures provides useful practical guidance.
3. Potential Impact
Autonomous driving: If latent communication proves robust in practice, it could fundamentally change V2V communication protocols. The 20× latency reduction over language-based methods and 40-90% bandwidth reduction over visual sharing are compelling for real-time safety-critical applications.
Multi-agent AI systems: The agent identity confusion phenomenon and the shallow-vs-deep analysis have broader implications for any multi-agent system built on transformer architectures. The insight that deep representations become identity-entangled could inform design choices in multi-robot coordination, distributed AI systems, and federated inference.
Practical deployment: The training-free property significantly lowers the barrier to adoption—fleet operators could potentially upgrade existing single-vehicle VLA systems to collaborative ones without retraining.
However, the paper doesn't address several practical concerns: heterogeneous model architectures between vehicles, privacy implications of sharing internal representations, adversarial robustness of KV cache communication, and regulatory considerations.
4. Timeliness & Relevance
The paper is highly timely. VLA models for driving are rapidly maturing (ORION, SimLingo, LMDrive all from 2024-2025), and the question of how to enable collaboration among these models is nascent. Language-based approaches like LangCoop appeared very recently (2025), making LACO's critique of language communication overhead and its latent alternative immediately relevant.
The connection to concurrent work on latent-space communication in LLMs (KV cache sharing, latent reasoning) positions this work at the intersection of two active research fronts.
5. Strengths & Limitations
Key Strengths:
Notable Limitations:
Overall Assessment
LACO presents a well-motivated and practically significant contribution to collaborative autonomous driving. The identification and analysis of agent identity confusion is the paper's most important intellectual contribution, while the complete framework demonstrates strong empirical results with meaningful efficiency gains. The work would benefit from larger-scale experiments, real-world considerations, and formal analysis, but it represents a solid first step toward latent-space V2V communication for driving.
Generated May 22, 2026
Comparison History (27)
Paper 1 presents a more fundamental and broadly applicable contribution: reframing recursive neural network inference as approximate inference over latent trajectories, with principled diagnostics that generalize across tasks. The stochastic exploration framework is novel, theoretically grounded, and training-free, with strong empirical gains (85.9%→98.0% on Sudoku). Its insights about inference-time computation apply broadly to recursive/iterative architectures. Paper 2, while practical, addresses a narrower application domain (collaborative driving) with more incremental engineering contributions combining known techniques. Paper 1's conceptual framework has greater potential to influence multiple research directions.
Paper 2 (LACO) has higher estimated impact due to its direct real-world relevance (connected autonomous driving), timeliness (foundation-model-era communication and latency constraints), and broad applicability of training-free latent communication/adaptation ideas across multi-agent robotics. Its focus on reducing communication/inference latency while maintaining performance targets a key deployment bottleneck. Paper 1 is novel and rigorously evaluated (including humans) but is demonstrated in a narrower domain (Overcooked) and is less immediately translatable to safety-critical applications than collaborative driving.
LACO addresses a practical and growing problem in collaborative autonomous driving with a well-motivated, multi-component solution (ILD, CHSA, SSKD) that tackles real engineering challenges of latency and information loss. It has broader applicability across the autonomous driving and multi-agent communication communities. Paper 1, while offering useful benchmark critique of ARC-AGI-3, is primarily a negative result showing benchmark weaknesses, with limited generalizability beyond this specific benchmark. The AERA agent's performance is modest, and the contributions are more diagnostic than constructive compared to LACO's actionable framework.
Paper 1 likely has higher scientific impact due to stronger novelty and broader relevance: it advances neurosymbolic learning for formal proof optimization with new structural metrics, an expert-iteration training pipeline, and demonstrated small-model gains over much larger baselines in Lean 4—useful for formal verification, automated theorem proving, and LM training data quality. Its methodological contribution (scaffolding + metrics + scalable optimization framing) can generalize beyond one domain. Paper 2 is timely and application-driven, but is narrower (collaborative driving in CARLA) and “training-free adaptation” may limit scientific depth and transfer compared with the foundational nature of Paper 1.
Paper 1 addresses fundamental, theoretical limitations of AI models (like LLMs) and introduces a predictable 'Deterministic Horizon' for model accuracy based on architecture. This offers profound implications and broad applicability across all AI subfields, fundamentally altering how models are designed and evaluated. In contrast, Paper 2 presents a specialized, applied solution for collaborative driving, which, while highly practical and useful for autonomous vehicles, has a narrower scope and less potential for widespread paradigm-shifting scientific impact.
Paper 1 likely has higher scientific impact because it introduces a concrete, novel technical framework (training-free latent communication with ILD/CHSA/SSKD) addressing clear bottlenecks (latency, information loss, identity confusion) and demonstrates closed-loop results in CARLA, supporting methodological rigor and near-term real-world relevance for connected/autonomous vehicles. Paper 2 is a broad survey/conceptual taxonomy; it is timely and could influence agendas across fields, but it contributes fewer directly testable methods or empirical advances, making its immediate scientific/engineering impact less certain.
Paper 2 addresses a critical and widespread challenge in AI: evaluating LLM agents for complex knowledge work. Its proposed benchmarking framework and guidelines have a broad impact across multiple domains, whereas Paper 1, though highly innovative, is restricted to the specific niche of collaborative autonomous driving. The breadth of impact and timeliness of evaluation standards for LLM agents give Paper 2 a higher potential scientific impact.
Paper 2 has higher likely scientific impact due to broader cross-domain applicability (any regulated AI system), strong timeliness with escalating AI governance requirements, and a rigorous, formal, executable framework (RDF/OWL, SHACL, PROV-O) that enables reproducible compliance validation. Its compiler/IR-to-modules approach and open-source artifacts increase adoption potential. Paper 1 is innovative for collaborative driving latency/communication, but its impact is narrower to multi-agent autonomous driving and relies on simulation evidence, limiting breadth relative to governance infrastructure.
Paper 2 addresses a critical healthcare problem—ventilator management in ICUs—with a novel human-in-the-loop multi-agent framework combining contextual bandits for preference learning with structured clinician feedback. Its direct clinical applicability, focus on safety/auditability (key barriers to AI adoption in medicine), and personalization approach give it broader real-world impact. Paper 1, while technically interesting in collaborative driving, proposes incremental improvements (training-free latent communication) in a narrower domain. Paper 2's cross-disciplinary relevance (AI, clinical medicine, HCI) and timeliness regarding safe AI deployment in high-stakes settings suggest greater impact potential.
TerminalWorld introduces a novel scalable benchmarking methodology that addresses a significant gap in evaluating AI agents on real-world terminal tasks. Its automated data engine processing 80K+ recordings, coverage of 1,280 unique commands across 18 categories, and demonstration that current frontier models achieve only 62.5% pass rate provide high-impact insights for the rapidly growing AI agent community. The benchmark fills a distinct niche (weak correlation with existing benchmarks) and is designed to scale with evolving practices. While LACO presents solid technical contributions to collaborative driving, its scope is narrower, constrained to a specific simulation environment, and builds incrementally on existing paradigms.