AI TrackerFrontier desk · AEST
Back to news
normalaitech

ARCHER: From Rationale Imitation to Executable Reasoning in Small Language Models

arXiv cs.AI · 14 Nov 2026, 04:05 AEST

ARCHER is a new framework designed to enhance small language models' capabilities in mathematical reasoning tasks. It employs a two-stage training process that focuses on producing executable and auditable reasoning trajectories.

Key facts

Read aloud

ARCHER is a new framework designed to enhance small language models' capabilities in mathematical reasoning tasks. It employs a two-stage training process that focuses on producing executable and auditable reasoning trajectories. Small language models are attractive for local, low-cost, and privacy-sensitive deployment, but they remain unreliable on mathematical reasoning tasks that require decomposition, exact computation, tool use, and verification. In the first stage, ARCHER compiles free-form chain-of-thought rationales into typed action traces containing reasoning steps, calculator calls, deterministic runtime observations, verification actions, and final answers. In the second stage, ARCHER refines this policy through asymmetric hierarchical reinforcement learning, using grouped and conditionally gated trajectory rewards for answer correctness, tool grounding, process discipline, and trace alignment. Across five standard sub-10B backbones, ARCHER improves over the strongest tool-augmented baseline by five point five one percentage points on macro-average across benchmark accuracies, including gains of eight point nine five points on MATH-five hundred and five point three three and five point three four points on AIME-twenty twenty four and AIME-twenty twenty five. These results demonstrate that executable supervision and reward-guided behavioral refinement provide a general and effective training paradigm for compact mathematical reasoners.

Sources in this cluster