SAM-G-Reasoning is a fine-tuned variant of SAM-G, trained on 196,000 verified multi-step reasoning traces and action plans. It is a 30.3M-parameter model designed for one specific capability: procedural reasoning on the edge, without any cloud dependency.
Where most small models attempt to be general assistants, SAM-G-Reasoning is deliberately narrow. It excels at decomposing instructions into ordered steps, tracking execution state across multiple operations, and emitting structured action plans. It is not a calculator, and it does not pretend to be one.
What it does
The model emits explicit reasoning traces in the form: step 1 ... step 2 ... Answer: X for questions, and ordered JSON plans for multi-step instructions. This transparency makes it auditable: you can inspect every step the model took before it reached its conclusion.
This design is particularly well suited to agentic control loops, where an agent must decompose a high-level instruction into a sequence of executable actions, track which steps have been completed, and adjust its plan when the environment changes.
Benchmark results
SAM-G-Reasoning was evaluated on twelve verified reasoning families. The aggregate exact-match score on a held-out evaluation set is 77.8%. The results reveal a clear pattern: the model is strong on procedural and state-tracking tasks, and limited on calculation-heavy tasks, which is expected at this scale.
| Reasoning Family | Exact Match | Type |
|---|---|---|
| Logic (modus ponens, tollens, chains) | 100% | Procedural |
| Plan chain (multi-step actions) | 100% | Procedural |
| Conversion (unit chains) | 100% | Procedural |
| Date and time (clock, calendar) | 92% | Procedural |
| Compare (max, min) | 92% | Procedural |
| State tracking (device toggles) | 83% | Working memory |
| Parity and digits | 58% | Mixed |
| Count and filter | 67% | Calculation |
| Sort list | 50% | Calculation |
| Word problem | 50% | Calculation |
| Arithmetic chain | 42% | Calculation |
State tracking at 83% is notable for a model of this scale. Maintaining a mutable state across several sequential operations requires the model to hold intermediate results in context and update them correctly with each step. The arithmetic and sorting families plateau because exact multi-digit calculation is not reliably learnable at 30M parameters. For those tasks, the recommended approach is to delegate to a deterministic tool and use SAM-G-Reasoning as the planning and orchestration layer.
Intended use
SAM-G-Reasoning is designed to serve as the planning and state-tracking layer of an agentic system. It decomposes a high-level instruction into an ordered sequence of steps, tracks execution state, and emits structured action plans. Arithmetic, data lookups, and other deterministic operations should be delegated to purpose-built tools.
The model runs entirely offline. It has no network dependency and can be deployed on embedded systems, edge servers, and robotic controllers with constrained resources.
Relationship to SAM-G
SAM-G-Reasoning is a fine-tuned specialisation of SAM-G. It inherits the base model's dual-mode architecture and on-device efficiency, and adds explicit reasoning trace generation. The two models are complementary: SAM-G handles general structured action planning, while SAM-G-Reasoning handles tasks that require step-by-step justification and state tracking.
