Atomic Fact Lookahead Llm Agent Planning Paper Poc 2026

Slug: atomic-fact-lookahead-llm-agent-planning-paper-poc-2026
Paper: arXiv:2506.09171 — "Improving LLM Agent Planning with In-Context Learning via Atomic Fact Augmentation and Lookahead Search"
Authors: Samuel Holt, Max Ruiz Luyten, Thomas Pouplin, Mihaela van der Schaar (University of Cambridge)
Venue: ICML 2025 — First Workshop on Computer Use Agents (Oral)
Date: 2026-05-19
Environment: macOS 14, Python 3.12 (stdlib only, no external packages)

Objective

Reproduce the core Atomic Fact Augmentation pattern from arXiv:2506.09171 on a minimal TextFrozenLake environment without any LLM API calls. Verify that fact accumulation across episodes leads to measurable performance improvement vs. a baseline with no memory.

Commands Run

python3 /tmp/atomic_fact_poc.py

Script: /tmp/atomic_fact_poc.py (self-contained, stdlib only)

Environment Setup

• Python 3.12.x (macOS)
• No external dependencies required
• Grid: 5×5 TextFrozenLake, goal at (4,4), 5 fixed holes at (0,3), (1,1), (2,4), (3,2), (4,1)
• Max steps per episode: 30
• Episodes: 10
• Random seed per episode: ep * 42

What Was Implemented

Three components matching the paper's LWM-Planner architecture:

1. Action Proposal (fact_agent_action): goal-directed movement with lookahead depth=1 filtering known holes
2. Atomic Fact Extraction (extract_facts): extracts {type, pos, text} records after each episode from trajectory data
3. Fact Deduplication: simple key-based dedup by (type, pos) to simulate the paper's predictive-consistency filter

Baseline agent uses identical goal-directed heuristic but never accumulates facts.

Output / Results

=== Atomic Fact Augmentation PoC (arXiv:2506.09171) ===
Environment: TextFrozenLake 5x5, 5 holes
Episodes: 10, Max steps/episode: 30

Baseline agent:       Success 0/10, avg steps 3.3
Fact-augmented agent: Success 5/10, avg steps 9.2

Facts accumulated by episode 10: 6 unique facts
Sample facts:
  - cell (2, 4) is a hole — action 'down' from (1, 4) leads to failure
  - cell (4, 1) is a hole — action 'down' from (3, 1) leads to failure
  - cell (1, 1) is a hole — action 'down' from (0, 1) leads to failure
  - cell (3, 2) is a hole — action 'right' from (3, 1) leads to failure
  - cell (4, 4) is the goal — reached via action 'right' from (4, 3)

Per-episode outcomes:
  Ep1: baseline=hole    (6 steps)  |  fact-agent=hole    (6 steps)
  Ep2: baseline=hole    (5 steps)  |  fact-agent=hole    (5 steps)
  Ep3: baseline=hole    (2 steps)  |  fact-agent=hole    (2 steps)
  Ep4: baseline=hole    (2 steps)  |  fact-agent=hole    (12 steps)
  Ep5: baseline=hole    (2 steps)  |  fact-agent=goal    (12 steps)
  Ep6: baseline=hole    (2 steps)  |  fact-agent=hole    (3 steps)
  Ep7: baseline=hole    (2 steps)  |  fact-agent=goal    (24 steps)
  Ep8: baseline=hole    (5 steps)  |  fact-agent=goal    (10 steps)
  Ep9: baseline=hole    (5 steps)  |  fact-agent=goal    (10 steps)
  Ep10: baseline=hole    (2 steps)  |  fact-agent=goal    (8 steps)

Key Observations

• Baseline: 0% success across all 10 episodes. Greedy goal-directed movement hits holes repeatedly with no learning.
• Fact-augmented agent: 50% success rate (5/10). First goal reached on Ep5 after accumulating 4 hole facts from Ep1–4.
• Late episodes (Ep8–10): consistent goal-reaching as the hole map became more complete.
• Ep4 divergence: fact-agent takes 12 steps vs baseline's 2 — the agent is deliberately taking longer routes to avoid known holes, trading step efficiency for survival. This matches the paper's observation that LWM-Planner takes longer paths but achieves higher task completion.
• Average steps: fact-agent takes 9.2 steps vs baseline's 3.3 — the higher count reflects safer paths, not inefficiency.

Limitations

• No real LLM calls: actions and fact extraction are deterministic heuristics simulating the LLM's role. The actual paper uses GPT-4 / Claude for action proposal, world model simulation, and state-value estimation.
• No lookahead depth >1: paper implements depth-limited tree search; this PoC uses depth=1 lookahead only.
• Fixed seed per episode reduces variance. In the paper, stochasticity is introduced by LLM sampling (temperature > 0 during action proposal, temperature = 0 during lookahead).
• No fact compression/summarization step (paper uses an optional LLM-based compressor for large fact sets).

Verdict

The core claim is reproducible: accumulating atomic facts across episodes measurably improves agent success rate (0% → 50%) without any weight updates. The pattern is directly applicable to LangGraph or Pydantic AI agents by injecting a fact store into the system prompt at each step.