ARCHER trains small language models to reason with calculator-grounded action traces
The two-stage training framework converts free-form chain-of-thought into typed, executable action traces and refines them with hierarchical reinforcement learning, reporting gains across sub-10B models on math benchmarks.
In plain words
ARCHER is a training method for small language models that teaches them to solve math problems by writing step-by-step plans the computer can check and run, including calculator calls and verification steps, instead of just imitating explanatory text. The authors report that models trained this way answer more benchmark math questions correctly than comparable tool-augmented baselines.
What was announced
The paper's authors argue that small language models are attractive for local, low-cost, and privacy-sensitive deployment but remain unreliable on math tasks requiring decomposition, exact computation, tool use, and verification. ARCHER addresses this in two stages. First, it compiles free-form chain-of-thought rationales into typed action traces containing reasoning steps, calculator calls, runtime observations, verification actions, and final answers, shifting supervision from imitating explanatory text toward a tool-grounded reasoning policy. Second, it refines that policy with asymmetric hierarchical reinforcement learning, using grouped and conditionally gated trajectory rewards for answer correctness, tool grounding, process discipline, and trace alignment.
On the results side, the authors report that across five standard sub-10B backbones, ARCHER improves over the strongest tool-augmented baseline by 5.51 percentage points on the macro-average of benchmark accuracies, including gains of 8.95 points on MATH-500 and 5.33 and 5.34 points on AIME-2024 and AIME-2025. Tests on two additional reasoning-specialized 7B backbones suggest executable grounding remains beneficial even after reasoning-focused reinforcement learning or distillation. The authors present this as evidence that executable supervision plus reward-guided refinement is a general training paradigm for compact mathematical reasoners; the figures are self-reported and not independently confirmed.
What it does not show
- Evidence is limited to the paper's abstract, so method details, experimental setup, and full results cannot be verified from the supplied material.
- All reported accuracy gains are the authors' own claims; no independent confirmation is available in the sources.
For security and IT teams
- Scope
- The paper targets small language models for local, low-cost, and privacy-sensitive deployment on mathematical reasoning tasks, per the abstract.
Sources
Source-support review completed 17 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Coverage: arXiv: abstract.
Updates and corrections
No corrections so far.
First published 17 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.