ProD: Progress-Aware Distillation Mitigates Stagnation in Small Language Model Agents Across 8 Architectures

Background and the Problem

Transferring agentic tool-using capabilities from frontier LLMs into small language models (SLMs, 0.6B to 8B parameters) via supervised fine-tuning (SFT) is critical for low-latency and on-device deployment. However, existing SFT distillation suffers from a pervasive operational breakdown: while small models memorize basic tool syntaxes, they struggle to monitor long-horizon execution progress. When an action yields an unexpected response or runtime error, SLMs frequently collapse into 'progress stagnation'—entering repetitive loops, re-executing identical invalid commands, and exhausting the context limit without making forward progress. In arXiv:2610.19439, researchers introduced ProD (Progress-Aware Distillation) to overcome this limitation.

Architecture and How It Works

1. Formalized Progress Stagnation Metrics:

• Systematically defines turn-wise and trajectory-level stagnation scores, evaluating whether sequential Reason-Act-Observe steps produce meaningful environmental state transitions;

1. Adaptive Margin Penalization via Stagnation Gaps:

• Inspired by direct preference alignment (SimPO) but moving beyond static offline margins, ProD calculates the real-time stagnation gap between student rollouts and non-stagnant teacher paths;
• Applies an adaptive margin penalty that dynamically penalizes unconstructive retries, compelling the student model to adopt the teacher's recovery and backtracking strategies;

1. Broad Model Scale Invariance:

• Formulated to operate seamlessly across SLMs spanning 0.6B ultra-lightweight models to standard 8B edge backbones.

Benchmarks and Measured Results

Evaluated within the CodeAct harness across 8 distinct open SLMs ranging from 0.6B to 8B parameters (Sections 3 & 4):

• Stagnation Reduction:
• While conventional SFT models exhibit stagnation rates exceeding 40% on complex extended tasks, ProD reduces progress stagnation by over 50% across the board;
• Task Success Amplification:
• By eliminating unconstructive dead-loops, agents achieve substantial double-digit percentage gains in task completion rates within standard step budgets;
• Generalizes reliably across both in-domain and out-of-domain evaluation environments;
• Efficiency and Token Conservation:
• Eliminates wasteful token loops, trimming average multi-turn token consumption by 25% to 35%.

Getting Started for Developers

• Paper Reference: Architectural specifics and training formulations are documented in arXiv:2610.19439;
• Engineering Recommendations: Developers fine-tuning small language models for local terminal assistants, embedded robots, or CLI copilots should avoid naive SFT on raw teacher logs. Adopting ProD's stagnation-penalizing objective ensures robust multi-turn recovery on resource-constrained hardware.