Open-S1-GRT-30M (MatterWaveAI/open-s1-grt-30m)

Open-S1-GRT-30M is a zero-nn.Linear System-1 Calibrated Agent Reflex & Decision Scorer built on the 29.31M-parameter MatterWaveAI/grt-30m Geometric Rotation Transformer backbone (12 blocks, d_model = 256, 2,048 context window, 151,669 Qwen 3 BPE vocabulary).

Instead of invoking a slow autoregressive frontier LLM for routine agent control-loop and verification decisions, open-s1-grt-30m scores discrete choices in a single forward pass with temperature-calibrated probabilities and option-order permutation invariance (0.000 position bias).

---

Supported System-1 Verification & Routing Domains

Domain	Purpose	Decision Output	Calibration Temperature
Action Routing	Select next agent step across tool, edit, test, search, ask, or finish	Up to 8 candidate actions	1.12
Complexity Escalation	Triage whether a query can be handled locally (simple) or escalated (complex)	simple, complex	1.08
Loop Guard	Detect unproductive agent tool loops vs. productive progress	continue, intervene, abort	1.05
Specification Support	Verify whether an agent claim or diff is entailed by the task specification	supported, unsupported, ambiguous	1.15
Verification Gate	Gate whether code changes are ready to commit or require testing/repair	pass, run_tests, repair	1.06
Ontological Hierarchy	Verify 5-tier conceptual hierarchy and prerequisite ordering	5 hierarchical tiers	1.04
Pedagogical Verification	4-mode multi-perspective conceptual verification	4 reasoning perspectives	1.05
Entity Disambiguation	Disambiguate surface mentions to canonical entities	linked, ambiguous, unlinked	1.07
Single-Hop Factual Verification	Verify 1-hop factual relation consistency	4-way candidate verification	1.03
Multi-Hop Chain Verification	Verify multi-hop (2–3 hop) relational chain consistency	valid_chain, invalid_hop, order_inverted, disconnected	1.06
Causal Science & Commonsense	Verify causal scientific and functional commonsense relations	4-way candidate verification	1.04

---

Benchmarks & Position-Invariance Guarantees

Benchmark / Gate	Metric	open-s1-grt-30m	Legacy nano (8.7M)	Chance
Paired-Shuffle Position Guard	Permutation Consistency	1.000	0.812	---
Paired-Shuffle Position Bias	Order Sensitivity	0.000	0.188	0.000
11-Domain System-1 Macro Accuracy	Held-Out Calibrated Eval	93.41%	---	27.6%
Held-Out Knowledge Transfer (100 tasks)	4-Way Accuracy	92.0%	18.0%	25.0%
Intel Arc XPU Accelerated Training	Wall-Clock (vs CPU)	101.19s (4.29x)	---	433.86s (CPU)

---

Complete MatterWaveAI Research Suite on Hugging Face

1. Base Causal LM: MatterWaveAI/grt-30m
2. Chat Fine-Tune: MatterWaveAI/grt-30m-chat
3. 4-Expert Mixture-of-Rotations (MoR): MatterWaveAI/grt-30m-mor
4. Circular-Pooled Sentence Embedding: MatterWaveAI/grt-30m-embedding
5. Cross-Encoder Reranker: MatterWaveAI/grt-30m-reranker
6. Vision-Language (GRTVL): MatterWaveAI/grt-30m-vl
7. Canonical Tokenizer & Polar Embedding Bank: MatterWaveAI/grt-30m-tokenizer
8. Rotational Text-to-Speech (GRTTTS): MatterWaveAI/grt-30m-tts
9. Multimodal Vision-Audio-Language-Speech (GRTOmni): MatterWaveAI/grt-30m-omni
10. System-1 Agent Reflex & Decision Scorer (open-s1): MatterWaveAI/open-s1-grt-30m
11. Rotational JEPA & Causal World-Model Core: MatterWaveAI/helmholtz-1
12. 5-Modality Platonic Omni-JEPA World Model: MatterWaveAI/helmholtz-1-omni
13. Rotational Selective State-Space (Mamba) World Model: MatterWaveAI/helmholtz-mamba-1
14. Unitary Gated Rotational RNN (minGRU / GRU) World Model: MatterWaveAI/helmholtz-rnn-1
15. System-1 Latent Trajectory Judge & MPC Router: MatterWaveAI/s1-grt-jepa
16. Streaming O(1) Rotational SSM World-Model Judge: MatterWaveAI/s1-grt-mamba
17. Stateful O(1) Rotational RNN World-Model Judge: MatterWaveAI/s1-grt-rnn

• Whitepaper: The Geometric Rotation Transformer (GRT): A Purely Rotational Foundation Architecture (GRT_Whitepaper.pdf)

---

License

Apache 2.0