When to ASK: Uncertainty-Gated Language Assistance for Reinforcement Learning

Juarez Monteiro; Nathan Gavenski; Gianlucca Zuin; Adriano Veloso

cs.AIApr 2, 2026

When to ASK: Uncertainty-Gated Language Assistance for Reinforcement Learning

Authors: Juarez Monteiro, Nathan Gavenski, Gianlucca Zuin, Adriano Veloso

Organizations: Kunumi Institute Belo Horizonte, MG, Brazil · King’s College London London, England, UK

Abstract

Reinforcement learning (RL) agents often struggle with out-of-distribution (OOD) scenarios, leading to high uncertainty and random behavior. While language models (LMs) contain valuable world knowledge, larger ones incur high computational costs, hindering real-time use, and exhibit limitations in autonomous planning. We introduce Adaptive Safety through Knowledge (ASK), which combines smaller LMs with trained RL policies to enhance OOD generalization without retraining. ASK employs Monte Carlo Dropout to assess uncertainty and queries the LM for action suggestions only when uncertainty exceeds a set threshold. This selective use preserves the efficiency of existing policies while leveraging the language model's reasoning in uncertain situations. In experiments on the FrozenLake environment, ASK shows no improvement in-domain, but demonstrates robust navigation in transfer tasks, achieving a reward of 0.95. Our findings indicate that effective neuro-symbolic integration requires careful orchestration rather than simple combination, highlighting the need for sufficient model scale and effective hybridization mechanisms for successful OOD generalization.

Figures & tables

		(a) Same-size evaluation				(b) Downward generalization (trained on 8×8)
Size	Model	Reward	Length	IR (%)	OR (%)	Reward	Length	IR (%)	OR (%)
$4\times 4$	PPO	–	–	–	–	$0.00\pm 0.00$	$46.98\pm 49.18$	–	–
	0.5B	–	–	–	–	$0.00\pm 0.00$	$41.12\pm 26.20$	100.00	61.71
	1.5B	–	–	–	–	$0.37\pm 0.49$	$6.92\pm 6.20$	100.00	42.98
	3B	–	–	–	–	$0.00\pm 0.00$	$44.58\pm 26.47$	100.00	64.52
	7B	–	–	–	–	$0.00\pm 0.00$	$61.00\pm 0.00$	100.00	95.57

TABLE I : Comparison of PPO baseline and ASK on FrozenLake: same-size evaluation and downward generalization.

Appendix figures & tables4 assets

Supplementary material from the paper’s appendix.

Appendix

Panel of Figure 3 — Fig. 2 : Schematic and image representations for the FrozenLake-v $1$ environment. In Figs. 2(d) and 2(b) , each cell corresponds to a Markovian state, and special tiles denote the start (S), goal (G), and terminal-failure states (H). These structured state descriptions are encoded as contextual information and incorporated into prompts, supporting decision-making in the PPO policy through gated interventions.

Model	Avg. $\tau$	Avg. Reward
Qwen2.5-0.5B	0.971	0.83
Qwen2.5-1.5B	0.544	0.83
Qwen2.5-3B	1.063	0.80
Qwen2.5-7B	0.944	0.82
Qwen2.5-14B	1.079	0.82
Qwen2.5-32B	0.763	0.83

Appendix

TABLE II : Best uncertainty threshold

\tau

per model (averaged across environment sizes).

Size	Model	Split	Reward	Length	IR (%)	OR (%)
$6\times 6$	PPO	Eval	$0.89\pm 0.31$	$9.22\pm 2.32$	–	–
	PPO	Test	$0.93\pm 0.26$	$9.49\pm 1.89$	–	–
	0.5B	Eval	$0.89\pm 0.31$	$9.22\pm 2.32$	21.10	0.02
	0.5B	Test	$0.93\pm 0.26$	$9.49\pm 1.90$	49.83	0.00
	1.5B	Eval	$0.89\pm 0.31$	$9.22\pm 2.32$	27.88	0.00
	1.5B	Test	$0.93\pm 0.26$	$9.49\pm 1.90$	55.03	0.00

Appendix

TABLE III : Comparison between PPO baseline and ASK on FrozenLake. Eval corresponds to validation performance using the gating mechanism. The test corresponds to in-domain test performance. Results are reported as mean

\pm

standard deviation over 100 episodes.

Env	LLM Model	Split	Reward	Length	IR (%)	OR (%)
$4\!\times\!4$	0.5B	Eval	$0.00\pm 0.00$	$51.76\pm 46.85$	100.00	57.13
	0.5B	Test	$0.00\pm 0.00$	$41.12\pm 26.20$	100.00	61.71
	1.5B	Eval	$0.47\pm 0.50$	$7.54\pm 6.46$	100.00	41.00
	1.5B	Test	$0.37\pm 0.49$	$6.92\pm 6.20$	100.00	42.98
	3B	Eval	$0.00\pm 0.00$	$83.40\pm 36.87$	100.00	62.20
	3B	Test	$0.00\pm 0.00$	$44.58\pm 26.47$	100.00	64.52

Appendix

TABLE IV : Downward generalization results. PPO trained on FrozenLake-8 and evaluated on smaller environments. Eval corresponds to validation performance, while Test reports final performance using fixed thresholds. Results are reported as mean

\pm

standard deviation over 100 episodes.

Uncertainty Estimation in RL

See all 24 papers

ASK in the Dark: Uncertainty-Gated LLM Assistance under Partial Observability
Jul 2, 2026Juarez Monteiro, Nathan Gavenski, Guilherme Lima +3Partially Observable RL LLM Prompting
Kunumi Institute, Belo Horizonte, MG, Brazil · King’s College London, London, United Kingdom
Smaller Abstract State Spaces Enable Cross-Scale Generalization in Reinforcement Learning
May 19, 2026Nasehatul Mustakim, Lucas LehnertPartially Observable RL OOD Generalization
Department of Computer Science University of Saskatchewan Saskatoon, Saskatchewan, Canada
Skill0.5: Joint Skill Internalization and Utilization for Out-of-Distribution Generalization in Agentic Reinforcement Learning
May 27, 2026Jiapeng Zhu, Jianxiang Yu, Yibo Zhao +5Agentic RL OOD Generalization
1East China Normal University, 2Meituan Longcat Team · 1East China Normal University · 2Meituan Longcat Team

Jul 2, 2026J/K move · Enter open · S save

ASK in the Dark: Uncertainty-Gated LLM Assistance under Partial Observability

Juarez Monteiro, Nathan Gavenski, Guilherme Lima, Francisco Galuppo+2

Kunumi Institute, Belo Horizonte, MG, Brazil · King’s College London, London, United Kingdom

Partially Observable RL LLM Prompting LLM-Guided RL Policy Learning

Open paper AskPDF arXiv

Reinforcement learning agents operating under partial observability must act on incomplete information, making them natural candidates for guidance from small language models (SLMs) that carry broad reasoning priors. Yet integrating SLM guidance into this setting has proven difficult: across all test environments, vanilla uncertainty-gated approaches achieve an overwrite rate at or near zero, meaning the SLM almost never contributes an independent action. We trace this failure to the bare egocentric prompt, which provides insufficient context for genuine reasoning, and identify it as a context problem rather than a capacity problem. We propose ASK+, which supplies the SLM with trajectory-aware context (a partially revealed map, visited positions, and action history) and structured chain-of-thought reasoning, converting it from a passive redundancy check into a more informative consultant that occasionally corrects the policy. We further establish that the predictive entropy signal used for selective querying measures action uncertainty rather than state uncertainty and remains informative in POMDPs, making uncertainty-gated assistance viable beyond fully observable settings. The stateful prompt drives substantial gains: on DoorKey, where vanilla ASK matches PPO (both 89%), ASK+ reaches 93% success; on FourRooms, success climbs from 53% to 70%; on HigherLower, accuracy reaches 73.7%, matching the SLM-only upper bound. Across all environments, Qwen3.5-2B matches or exceeds Qwen3.5-4B, confirming that prompt design and selective gating dominate the impact of model scale, enabling guidance without large models.

Abstract

Uncertainty Estimation in RL

Explore similar work

ASK in the Dark: Uncertainty-Gated LLM Assistance under Partial Observability

Smaller Abstract State Spaces Enable Cross-Scale Generalization in Reinforcement Learning

Skill0.5: Joint Skill Internalization and Utilization for Out-of-Distribution Generalization in Agentic Reinforcement Learning