Organizations: Hangzhou Institute for Advanced Study, University of Chinese Academy of Sciences, Hangzhou, China · Zhejiang University, Hangzhou, China
Large Language Models (LLMs) are widely adopted for reflective error correction in task planning; however, existing approaches primarily rely on autonomous exploration, rendering them prone to accumulating erroneous experiences during early interactions with new environments. To address this limitation, we propose the Flexible Instructional Scaffolding Reflection Framework (FISRF), an agent--mentor architecture that estimates the robot's Zone of Proximal Development (ZPD) in task planning and adaptively selects reflection strategies and mentor guidance, thereby facilitating the accumulation of effective error-correction experience. Experiments in ALFWorld and on real-world robotic platforms demonstrate that FISRF reduces the average number of error-correction rounds by 37% and redundant reasoning by 25.1%, while improving the overall task success rate by 8%. Project webpage : https://fisrfplanning.github.io/
Figures & tables
Fig. 1: Compared with existing approaches, FISRF adaptively determines whether to solicit mentor guidance and selects appropriate reflection strategies according to the difficulty of correction and the magnitude of improvement, in each trial, thereby effectively preserving and leveraging valuable experience for subsequent correction.
Fig. 2: The framework of FISRF. The planning process is executed by the Actor, the evaluation process is executed by Evaluator, the reflection process is executed by the Self-Reflection, the difficulty of correction and magnitude of improvement are assessed, according to the assessment results, determine whether to seek mentor guidance and implement appropriate reflection strategies. By integrating insights from autonomous exploration with mentor guidance, our framework inform subsequent plan. The resulting reflections and planning trajectories are stored in a memory module for long-term retrieval.
Fig. 3: The detailed illustration of FISRF. The difficulty of correction and magnitude of improvement are first assessed, determining reflection strategies and whether to seek mentor guidance. After that, the detailed self-reflection process consists of the valuable experience summary process, the failure lesson summary process, Seek mentor guidance and the Reflection Summary process. Final output of the self-reflection process is a improved new plan for the next attempt.
Fig. 4: In an ALFWorld example, FISRF corrects a failed trajectory by summarizing experiences in the first episode, subsequently learning from failures and seeking mentor guidance in the second episode.
Method
Success Rate (%) ↑
Put
Clean
Heat
Cool
Examine
SR@1
SR@2
Overall
Model: Gemini 2.5
Plan-Only
82.9
71.0
82.6
90.5
66.7
73.9
76.1
79.1
Reason-Only
90.2
80.6
79.2
95.7
55.6
76.1
82.1
83.6
Reason-Reflect
97.6
80.6
78.3
90.5
66.7
76.1
81.3
85.1
Reason-Flex Reflect
100.0
87.1
60.9
90.5
88.9
78.4
85.8
87.3
TABLE I: Success rates (%) of FISRF and baseline methods on ALFWorld across different task categories. Results are reported under Gemini 2.5 and GPT-4o model
Methods
Efficiency
Flexibility
Costref @4 ↓
Raterev(%) @4 ↑
Roundrev @2 ↓
Roundrev @4 ↓
R ↑
CV(%)↑
Model:Gemini2.5
Reason-Reflect
913.7
64.3
3.26
4.79
207
23.5
Reason-Flex Reflect
558.3
69.6
2.30
3.15
237
21.5
FISRF(Ours)
370.5
87.5
1.45
1.71
253
36.7
Model:GPT4o
TABLE II: Flexibility and Efficiency of reflection methods. Our FISRF demonstrates significantly higher flexibility and efficiency.
Fig. 5: Food collection task of a failed task successfully refined using our FISRF.
Fig. 6: Result of success rate and round revision with different episodes. Our FISRF outperforms in all episodes, converges more quickly and has a significant better efficiency.
Method
SR
R
CV
Roundrev
Costref
Raterev
(%)
( ↑ )
( ↑ )
( ↓ )
( ↓ )
( ↑ )
Full
94.8
253
36.7
1.7
370.5
87.5
w/o Exp
87.3
198
24.5
3.6
487.3
68.7
w/o Lesson
88.8
208
22.1
2.4
472.5
75.3
w/o Mentor
82.9
272
27.6
4.2
550.7
63.5
TABLE III: Ablation study of FISRF components in AlfWorld tasks. All designed modules contribute to the performance.
Large Reasoning Language Models (LRLMs) enable multi-step reasoning for robotic task planning, but continued reasoning can overwrite valid intermediate plans or leave constraint violations unresolved, reducing planning reliability and wasting inference-time computation. We develop an inference-time monitor that exposes and verifies intermediate plans without disrupting the original decoding trajectory. Building on this monitor, we propose SafeInferCom, a formal verifier-guided framework that preserves valid intermediate plans and directs error correction during generation. Experiments across multiple LRLMs and planning domains reveal reasoning-response inconsistency and limited self-correction under one-shot inference. SafeInferCom improves planning success and accelerates error correction relative to one-shot inference. When combined with iterative refinement, it further improves success while reducing token usage compared with refinement alone. We additionally evaluate SafeInferCom in VirtualHome and provide a real-world robotic-arm demonstration.
Weizhe Xu, Jialiang Fan, Mengyu Liu +1
University of Notre Dame · Washington State University
Embodied LLMs endow robots with high-level task reasoning, but they cannot reflect on what went wrong or why, turning deployment into a sequence of independent trials where mistakes repeat rather than accumulate into experience. Drawing upon human reflective practitioners, we introduce Reflective Test-Time Planning, which integrates two modes of reflection: reflection-in-action, where the agent uses test-time scaling to generate and score multiple candidate actions using internal reflections before execution; and reflection-on-action, which uses test-time training to update both its internal reflection model and its action policy based on external reflections after execution. We also include retrospective reflection, allowing the agent to re-evaluate earlier decisions and perform model updates with hindsight for proper long-horizon credit assignment. Experiments on our newly-designed Long-Horizon Household benchmark and MuJoCo Cupboard Fitting benchmark show significant gains over baseline models, with zero-shot generalization to photorealistic HM3D environments and real-robot experiments on a Franka Panda arm. Ablations confirm that reflection-in-action and reflection-on-action are mutually dependent, and that retrospective reflection achieves better credit assignment than step-wise external feedback at lower computational overhead. Qualitative analyses further highlight behavioral correction through reflection.
Large Language Models (LLMs) have enabled more natural human-robot interaction, but open-source models often exhibit unstable long-horizon reasoning and inefficient action execution when deployed in agentic robotic frameworks. This paper presents an enhanced ROS-Agent based architecture that improves task reliability and execution efficiency for agentic robotic systems using open-source LLMs. The proposed system introduces a novel intermediate mechanism, termed the MetaTool, which enforces structured planning prior to action execution. Given a natural-language command, the MetaTool induces the LLM to generate a pseudo-code plan of intended tool invocations, which is stored in the ROS-Agent's scratchpad and persists throughout execution. By explicitly separating planning from execution, the proposed approach reduces execution loops and improves deterministic behavior. The architecture is validated on a custom mobile robotic platform with multimodal perception and motion control capabilities. Experimental results on real-world interactive tasks demonstrate improved task completion and contextual consistency, with up to ~24% gains on complex tasks compared to the baseline framework.
Kazi Abrar Mahmud, Nilotpaul Kundu Dhurubo, Tamal Kirttonia +2
Department of Electrical and Electronic Engineering, Bangladesh University of Engineering and Technology (BUET), Dhaka 1000, Bangladesh · University of Massachusetts Amherst, 300 Massachusetts Ave, Amherst, MA 01003, United States