🤖 AI

CockpitHAT: Dependency-Graph-Driven Hierarchical Attribution for Embodied Multi-Agent Cockpits

The paper introduces CockpitHAT, a hierarchical attribution framework that leverages dependency graphs, multi-channel evidence, and safety-aware evaluation to effectively diagnose process-level failures in embodied multi-agent cockpit systems, outperforming existing text-only methods on both public and newly released benchmarks.

Original authors: Wei Wang, Shuanghe Liu, Zhu Zhuo, Jiaqi Zhong, Xiaozhao Zhao, Xiaojie Zuo, Jie Su

Published 2026-08-04

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Wei Wang, Shuanghe Liu, Zhu Zhuo, Jiaqi Zhong, Xiaozhao Zhao, Xiaojie Zuo, Jie Su

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are watching a team of robots work together to drive a car. They talk to each other, check the map, and decide when to turn the steering wheel. This is the world of multi-agent systems: groups of AI "agents" collaborating to solve problems. But here's the tricky part: sometimes the robots get the final destination right, but they took a terrible, dangerous path to get there. Maybe they almost crashed into a wall just to avoid a puddle, or they turned off the headlights because they misunderstood a joke. This is called "Correctness Collapse." It's like a student getting an 'A' on a math test because they guessed the right answer, but their work shows they forgot how to add numbers. In a car, guessing the right answer isn't enough; the process has to be safe, too.

To fix this, scientists need to play detective. They need to look at the robots' conversation and figure out exactly who made the mistake and when it happened. This is called attribution. Usually, detectives just read the chat logs. But in a car, the robots also talk to the engine, the GPS, and the sensors. If a robot says "I'm fine" but the engine is on fire, a chat log alone won't tell you the truth. You need to look at the whole picture: the words, the car's state, and the environment. This paper introduces a new way to be a detective that looks at all those clues at once, specifically for cars that drive themselves or have smart assistants.

The Detective's New Toolkit: CockpitHAT

The researchers from ByteDance built a new system called CockpitHAT to solve the "Correctness Collapse" problem in car cockpits. They realized that old detective methods were too simple. They usually looked at a conversation like a line of text, assuming that what happened right before a mistake was the cause. But in a complex team of robots, the cause of a crash might have happened ten steps ago, or it might be hidden in a signal the car sent to its brakes, not in the words the robots spoke.

The Dependency Map: Following the Trail, Not the Clock
Imagine you are trying to figure out who spilled a glass of milk. If you just look at who was standing next to the table when it fell, you might blame the wrong person. Maybe the person who knocked it over did it five minutes ago, and the person standing there now just happened to be there.

CockpitHAT uses a Dependency Graph instead of a simple timeline. Think of this as a map of "cause-and-effect" connections rather than a clock. It asks: "Did this action depend on that previous action?" If Robot A sent a command that Robot B used, they are connected, even if they spoke at different times. If two robots were just chatting about the weather but didn't affect the car's brakes, they are far apart on the map, even if they spoke one second apart. By slicing the investigation based on these connections (how "close" the actions are on the map) rather than just time, CockpitHAT can spot the real culprit much better.

The "Embodied" Adapter: Listening to the Car's Heartbeat
The biggest problem with old methods is that they only listen to the robots' voices. But in a car, the "voice" isn't just words; it's also the speedometer, the engine temperature, and the memory of what happened in the last drive.

CockpitHAT adds a special Embodied Channel Adapter. Imagine this as a translator who listens not just to the conversation, but also to the car's dashboard lights and the driver's nervousness. If a robot says "All clear" but the car's sensors detect a pedestrian, this adapter flags it immediately. It pulls that dangerous signal up to the front of the investigation, ensuring that safety warnings aren't buried deep in the history. It treats safety violations as the most important clues, forcing them to the front of the line regardless of when they happened.

The Safety-First Jury
Once the clues are gathered, CockpitHAT doesn't just pick one answer. It uses a panel of four AI analysts to vote on what happened. One of these analysts is a dedicated "Safety Expert" whose only job is to look for danger. If the group is arguing about whether a mistake was minor or life-threatening, the Safety Expert gets extra weight.

The system uses a special voting rule: if there is any chance the mistake could be dangerous (like a potential crash), the system assumes the worst-case scenario to be safe. It's better to be overly cautious and say "This is dangerous" than to miss a real threat. This ensures that the final report highlights the most critical failures, not just the most obvious ones.

What They Found

The team tested their new detective system on two types of challenges. First, they used existing tests (called Who&When) that had manually crafted and algorithmically generated failure traces. On these tests, CockpitHAT was a huge improvement over the previous best methods.

It correctly identified the wrong robot 77.9% of the time on hand-made tests and 86.5% on computer-generated tests.
It pinpointed the exact moment of the mistake 37.8% of the time on hand-made tests and 46.0% on computer-generated tests.
This was a massive jump of up to 17.6 points in accuracy compared to the old "text-only" champions.

Second, they created their own new test called COCKPITBENCH. This was a special collection of 212 failure stories specifically for cars, including real-world data and tricky edge cases. These stories were labeled with ISO 26262 ASIL severity levels (a standard way to rate how dangerous a failure is, from A to D).

On this tough new test, CockpitHAT found the wrong agent 78.3% of the time and the exact step 38.2% of the time.
It was particularly good at spotting safety violations, which are the most critical to catch.

Why It Matters

The paper suggests that for AI to be safe in real-world, physical environments like cars, we can't just look at the text they write. We have to understand how their actions depend on each other and how they interact with the physical world. CockpitHAT shows that by mapping these dependencies and listening to the car's sensors, we can catch dangerous mistakes that other systems miss.

However, the authors are careful to note that this is a tool for diagnosis after the fact, not a system that stops crashes in real-time. It helps engineers understand what went wrong so they can fix the robots before they drive again. They also admit that their system relies on clear signals and might struggle if the robots are using hidden knowledge or if the car is in a completely new, confusing situation. But for now, it offers a much clearer window into the "black box" of multi-agent car systems, turning a confusing mess of errors into a solvable mystery.

Technical Summary: CockpitHAT

Problem Statement

Large Language Model (LLM) multi-agent systems in safety-critical embodied environments, such as automotive cockpits, suffer from a phenomenon termed Correctness Collapse. In this scenario, systems may achieve high task-level accuracy while masking severe process-level failures. In automotive contexts, this is particularly hazardous because lexically correct utterances can trigger dangerous physical operations (e.g., disabling visibility lights) that are only detectable through non-textual channels like vehicle state or environmental context.

Existing attribution methods fail to address this regime for two primary reasons:

Methodological Limitations: Current approaches rely exclusively on text traces and positional windows, ignoring dependency structures and multi-channel evidence (vehicle state, environment, memory). They lack mechanisms to incorporate ISO 26262 safety severity labels (ASIL).
Benchmark Deficiencies: Existing evaluation suites are text-only, single-agent, or lack safety-severity annotations, systematically underrepresenting rare catastrophic failures and omitting critical safety labels.

Methodology: COCKPITHAT

The authors introduce COCKPITHAT, a hierarchical attribution framework designed to diagnose failures in embodied multi-agent systems by integrating dependency graphs, multi-channel evidence, and safety-aware consensus.

1. Dependency-Graph-Driven Hierarchical Slicing

Instead of using fixed positional windows, COCKPITHAT constructs a Directed Acyclic Graph (DAG) from the interaction trace.

Graph Construction: Nodes represent agent actions (text or tool calls). Directed edges are induced based on four operational signals: reference (entity linking), plan (execution of proposed plans), tool (tool output usage), and control (task delegation).
Dependency Distance: Context selection is based on the shortest path distance in the DAG ($dd$) rather than temporal adjacency.
Four-Layer Slicing: For any focal step, the context is sliced into four layers based on dependency distance:
- Immediate ($dd=1$): Verbatim content with full reasoning chains.
- Local ($dd=2$): Compressed to key actions and commitments.
- Distant ($dd=3,4$): Summarized outcomes (max 20 words).
- Global ( $dd \ge 5$ ): Sparse milestones (decisions, failures, hand-offs).

2. Embodied Channel Adapter (ECA)

To address failures invisible in text (e.g., safety violations, environmental misgrounding), an Embodied Channel Adapter is employed when the trace schema includes side channels (environment, vehicle state, memory).

Salience Scanning: A lightweight LLM scans the trace to flag steps associated with four failure classes: Interaction Risk, Environmental Misgrounding, Safety Violation, and Cross-Domain Contamination.
Channel-Aware Promotion: Flagged nodes are promoted to higher context layers (closer to the immediate layer) regardless of their dependency distance. Specifically, Safety Violations are forced into the immediate layer ( $L_1$ ).
Evidence Attachment: Relevant side-channel data (e.g., vehicle state $\sigma$ , environment $e$ , memory $m$ ) is attached to the promoted nodes based on the failure tag.

3. Safety-Uplifted Objective Consensus

The framework employs a panel of specialized analysts to resolve attribution, moving beyond symmetric cost aggregation.

Multi-Analyst Panel: Each trace is analyzed by $K=4$ agents: one mandatory Safety Analyst (prioritizing ISO 26262 severity) and three generic analysts sampled from a diverse persona pool.
Two-Stage Consensus:
1. Maximum-Severity Prior: The final severity rating ( $\eta^*$ ) is determined by the maximum predicted severity among confident analysts, preventing high-risk alarms from being out-voted.
2. Safety-Uplifted Aggregation: Analyst weights are adjusted asymmetrically based on the predicted severity. High-severity predictions receive a "safety uplift" (super-linear growth), ensuring that safety-critical failures are not diluted by generic consensus.
Disagreement Resolution: If marginal scores fall below a calibrated threshold, a protocol forces analysts to cite explicit evidence, with a mechanical tie-break favoring hypotheses supported by concrete trace evidence.

Key Contributions

COCKPITHAT Framework: A novel attribution system that replaces positional windows with dependency-distance thresholds, integrates multi-channel embodied evidence via an adapter, and applies a safety-uplift mechanism to consensus voting.
COCKPITBENCH: A new benchmark comprising 212 annotated failure traces spanning dialogue, vehicle-state, environmental, and memory channels.
- Traces cover eight failure categories and four ISO 26262 ASIL severity levels.
- Annotations are derived from a three-expert consensus process.
- The dataset includes hand-crafted, algorithmically generated, and real-world telemetry scenarios.

Experimental Results

The authors evaluated COCKPITHAT on the Who&When benchmark (Hand-Crafted and Algorithm-Generated splits) and their own COCKPITBENCH.

Performance on Who&When:
- Agent-Level Accuracy: Achieved 77.9% (Hand-Crafted) and 86.5% (Algorithm-Generated), surpassing the text-only SOTA (ECHO) by up to 17.6 and 16.7 points, respectively.
- Step-Exact Accuracy: Achieved 37.8% (Hand-Crafted) and 46.0% (Algorithm-Generated), outperforming ECHO by up to 16.7 points.
- Tolerance: The model showed significant gains in step localization even with tolerance (e.g., +12.3 points at $\pm 5$ steps on Hand-Crafted).
Performance on COCKPITBENCH:
- Achieved 78.3% agent-level and 38.2% step-exact accuracy.
- Ablation Studies:
  - Dependency Hierarchical Slicing was the dominant contributor to agent-level accuracy gains, delivering a +25.2% point jump on COCKPITBENCH.
  - Embodied Channel Adapter (ECA) was the decisive factor for step-level accuracy on COCKPITBENCH, improving step accuracy by +6.2% points (from 0.320 to 0.382) and agent accuracy by an additional +4.5% points. This component recovered error-onset steps that earlier text-only stages tended to compress.
  - Safety-Uplifted Consensus contributed to overall performance, with high-severity (ASIL-C/D) slices showing higher step accuracy, consistent with the ECA design that forces safety violations into immediate layers.

Significance and Claims

The paper claims that dependency-aware, multi-channel, risk-calibrated attribution is an effective paradigm for reliable failure diagnosis in real-world embodied LLM multi-agent systems.

Paradigm Shift: It moves attribution from outcome-based scoring to systematic error localization that accounts for physical state changes and safety severity.
Safety Integration: By coupling ISO 26262 ASIL labels with attribution, the framework ensures that high-risk failures are not masked by high task-level success rates.
Benchmarking: COCKPITBENCH addresses the structural limits of prior benchmarks by providing the first dataset with joint ISO 26262 severity labels and multi-channel embodied evidence.

The authors position this work as a necessary step toward diagnosing "Correctness Collapse" in safety-critical domains, noting that while the system is designed for post-hoc diagnosis, it establishes the foundation for future real-time safety interventions.