Research map / Reinforcement learning / Visual and representation RL

Zero-Shot Instruction Following in RL via Structured LTL Representations

NeurIPS 2026 · Reinforcement learning › Visual and representation RL · approach: In-context and adaptive RL · 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 LTL-GNN, which conditions a multi-task RL policy on sequences of Boolean formulae extracted from accepting runs of LDBAs and encoded by a GNN+GRU, evaluated zero-shot on ChessWorld.

Problem

Zero-shot instruction following of LTL specifications in multi-task RL is hard when multiple atomic propositions can hold simultaneously and interact, as existing automaton-conditioned methods treat assignments in isolation.

Approach

Translate automaton transitions into compact Boolean formulae (positive and negative conditions), encode their syntax trees with a GCN, and aggregate over the accepting run with a GRU to condition a PPO policy, so the policy can compose semantics of Boolean operators to generalize to unseen assignments.

Key results

• On finite-horizon task set phi_3, LTL-GNN reaches 82.6% success vs 68.8% for DeepLTL and 51.7% for Transformer.
• On infinite-horizon task set phi^inf_1, LTL-GNN achieves 86.0% success vs 40.0% (DeepLTL) and 43.1% (Transformer).
• LTL-GNN is best or tied on most finite-horizon sets, but DeepLTL is marginally better on phi_7 (91.9 vs 91.0) and phi^inf_GF (95.7 vs 92.8).

Surprising findings

• The Transformer-augmented DeepLTL baseline underperforms plain DeepLTL on most task sets, despite attention theoretically capturing assignment interactions.

Limitations

• Evaluated only in a single chess-based gridworld with five propositions and discrete actions.
• Relies on a known labeling function mapping states to atomic propositions.
• Boolean formula extraction uses an approximate template-based procedure with DNF fallback, not guaranteed minimal.

Experimental setup

Datasets: ChessWorld (procedurally generated LTL tasks)

Benchmarks: ChessWorld finite-horizon tasks phi_1-phi_7 ChessWorld infinite-horizon tasks ChessWorld reach-avoid difficulty sweep

Metrics: success rate average discounted return

Baselines: DeepLTL Transformer-augmented DeepLTL

Builds on: DeepLTL LTL2Action PPO GCN limit-deterministic Büchi automata

Closest papers by meaning

• DeepLTL: Learning to Efficiently Satisfy Complex LTL Specifications for Multi-Task RL (ICLR 2025 · similarity 0.92)
• One Subgoal at a Time: Zero-Shot Generalization to Arbitrary Linear Temporal Logic Requirements in Multi-Task Reinforcement Learning (NeurIPS 2025 · similarity 0.91)
• Foundation Policies with Hilbert Representations (ICML 2024 · similarity 0.87)
• Automaton Constrained Q-Learning (NeurIPS 2025 · similarity 0.87)
• Zero-Shot Reinforcement Learning via Function Encoders (ICML 2024 · similarity 0.87)
• Skill Machines: Temporal Logic Skill Composition in Reinforcement Learning (ICLR 2024 · similarity 0.87)