Research map / Multi-agent systems and games / Human-AI and goal-conditioned

Attention Trajectories as a Diagnostic Axis for Deep Reinforcement Learning

NeurIPS 2026 · Multi-agent systems and games › Human-AI and goal-conditioned · approach: Language-grounded reinforcement agents · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes a hierarchical-attention profile (h-profile) that aggregates LRP saliency over annotated objects/modalities to track attention trajectories during DRL training, linking them to behavior in Atari, custom Pong, and biomechanical visuomotor tasks.

Problem

Feature reliance in deep RL agents and how it emerges during training is poorly understood; saliency maps are usually applied qualitatively to isolated frames of trained agents, which invites post-hoc storytelling.

Approach

Aggregate pixel- or feature-level saliency (mainly LRP over top neurons of the penultimate layer) into object- or modality-level hierarchical attention profiles, computed on a fixed dataset across training, compared across conditions, validated with behavioral tests, and replicated with other saliency methods.

Key results

• ANOSIM showed significant algorithm-based separation of h-profiles in all four Atari games (R from 0.80 to 0.93, p=0.01).
• In Breakout, DQN and QR-DQN reallocated attention to bricks (~60%) and suffered severe degradation under brick color perturbation (Delta r approaching -1), while A2C and PPO stayed robust (Delta r ~0).
• In the moving-button biomechanical task, agents attended more to the buttons (paired t-test p=0.009, Cohen's d=-1.31), but had much lower scores (-4.5 vs 76.4).

Surprising findings

• Agents with the same reward trajectory differed systematically in attention; DQN and QR-DQN over-relied on bricks despite similar scores.
• In Distractor Pong, agents kept attending to the reward-irrelevant second ball and interacted with it in behavioral tests.

Limitations

• Relies on predefined object or modality labels, which limits applicability to raw pixel inputs.
• Object-level abstraction can bias the representation and miss relational features such as distances or velocities.
• Small number of seeds in the biomechanical case study (3 agents per task in some analyses).

Experimental setup

Datasets: Atari 2600 Breakout Atari 2600 Pong Atari 2600 Space Invaders Atari 2600 Ms. Pacman Custom Pong variants User-in-the-Box choice reaction and car parking tasks

Benchmarks: Atari 2600 Breakout Atari 2600 Pong Atari 2600 Space Invaders Atari 2600 Ms. Pacman Custom Pong User-in-the-Box choice reaction User-in-the-Box parking a remote control car

Metrics: h-profile attention share ANOSIM R statistic relative reward change under perturbation (Delta r) choice frequency in dual-ball test paired t-test p-value and Cohen's d episode score

Baselines: Grad SmoothGrad perturbation-based saliency

Builds on: Layer-wise Relevance Propagation SmoothGrad Stable-Baselines 3 User-in-the-Box Concept relevance propagation

Closest papers by meaning

• Visualizing and Understanding Atari Agents (ICML 2018 · similarity 0.90)
• Machine versus Human Attention in Deep Reinforcement Learning Tasks (NeurIPS 2021 · similarity 0.89)
• Estimating cognitive biases with attention-aware inverse planning (NeurIPS 2025 · similarity 0.86)
• Seek Commonality but Preserve Differences: Dissected Dynamics Modeling for Multi-modal Visual RL (NeurIPS 2024 · similarity 0.85)
• DrM: Mastering Visual Reinforcement Learning through Dormant Ratio Minimization (ICLR 2024 · similarity 0.84)
• Deep Reinforcement Learning with Time-Scale Invariant Memory (AAAI 2025 · similarity 0.84)