GS1-0.8B (General System One)

Fast, Faithful, Frontier System One Intelligence

GS1 is a state-of-the-art family of fast decision models developed by BIGAI (Beijing Institute for General Artificial Intelligence) designed to bring capable, uncertainty-aware decisions into the execution path of agentic intelligence. Rather than relying on traditional autoregressive text generation or fixed-label output projections, GS1 computes native probabilistic decision distributions directly over variable candidate sets.

---

📌 Model Details

• Developed by: BIGAI (Beijing Institute for General Artificial Intelligence)
• Correspondence: Zilong Zheng (zlzheng@bigai.ai)
• Model Family: GS1-0.8B, GS1-2B, GS1-4B, GS1-9B
• Base Architecture: Initialized from Qwen3.5 with hybrid Gated DeltaNet (GDN) and full-attention layers.
• License: Apache 2.0
• Primary Task: Agentic Decision Making, Routing, Verification, Extraction, and Control.

---

🔥 GS1 Highlights

GS1 systematically redesigned modern LLM components for decision-native execution:

• 🎯 Native Cross-Attention Decision Head: Instead of generating text or classification logits, GS1 queries evidence across candidates using multi-head cross-attention. This design guarantees exact head-level permutation-identical probabilities when candidate order is restored and context is held fixed.
• 🔀 Segmented Bidirectional Visibility: Combines causal Gated DeltaNet (GDN) layers with segmented bidirectional attention at full-attention (FA) sites. Candidates observe their full comparison context while keeping shared prefixes completely immutable and reusable across questions.
• ⛓️ Carry-Preserving Attention Residuals (AttnRes): Implements a depth-routing mechanism over sparse receivers to dynamically query earlier representations without replacing the explicit skip-connection anchor.
• 🏷️ Typed Native Output Space: Directly computes three unified decision modes from one joint distribution:
  • Choice: Dynamic multi-class probability distribution.
  • Noul: Binary true/false verification ($\pi(\text{true})$).
  • Score: Expected value calculated across ordinal level legends ($\sum k\pi_k$).

---

💡 Advanced Post-Training & Reliability

• ⚖️ Calibrated Reward Policy Optimization (CRPO): Adapts Reinforcement Learning with Calibration Rewards (RLCR) by using the exact sampled decision probability within a Brier-shaped penalty. This directly ties RL training signals to the output policy exposed at deployment.
• 🛑 Learned Abstention ("I Don't Know"): Jointly optimizes accuracy and non-hallucination rates (NHR) under Lagrangian constraints, turning rejection/abstention into an explicit, optimal action.
• 🎮 Multi-Step RL for Long-Horizon Control: Extends CRPO to interaction trajectories with shared terminal feedback. GS1 demonstrates strong sequential decision-making capabilities, including defeating the Very Hard StarCraft II computer opponent.
• 🔄 Dual-Loop Recursive Self-Improvement (RSI): Uses AI-driven loops for continuous data synthesis (evolving public world generators and executable oracles) and infrastructure tuning.

---

🏆 Performance Benchmarks

GS1 models establish a new Parameter-Performance Pareto Frontier on the Decision Index 0.4.1 public text benchmark across all evaluated parameter scales:

Model Scale	GS1 Score	Top Displayed Baseline Leader	Score Gain	Benchmark Pareto Status
GS1-0.8B	56.15	53.79 (StartLux v1.1 0.8B)	+2.36	🥇 Frontier Leader
GS1-2B	62.22	60.64 (GLIDE 2B)	+1.58	🥇 Exceeds Jev 1.14 reference (60.70)
GS1-4B	68.95	67.81 (StartLux v1.1 4B)	+1.14	🥇 Exceeds Jev 1.14 reference by +8.25
GS1-9B	70.47	69.98 (StartLux v1.1 9B)	+0.49	🥇 Overall Top Score

---

⚡ Inference & Computational Efficiency

Eliminating autoregressive text generation drastically shifts latency bounds to single-pass state encoding and head evaluation.

• ⏱️ Ultra-Low Latency: Warm engine execution on a single NVIDIA RTX 4090 (BF16 computation) achieves single-question decisions in 16.00 ms (3.38x speedup over single-reference baseline).
• 📊 High Batch Scaling: Achieves up to 4.72x speedup for 32-question workloads (reduced from 1689.74 ms to 357.76 ms) via CUDA graph optimization.
• 💾 State Reuse & KV Caching: Architecture-level support allows shared state KV cache and head cross-attention projections to be computed once and reused across multiple decision branches.

---

💻 Quick Start & Input Format

GS1 accepts structured XML payloads separating shared evidence, decision instructions, and options:

<state>
{shared state, observations, and context}
</state>
<question>
{decision instruction or policy query}
</question>
<option>
candidate_1_id: candidate_1_description
</option>
<option>
candidate_K_id: candidate_K_description
</option>

Citation

@techreport{zheng2026gs1,
  title={GS1: Fast, Faithful, Frontier System One Intelligence},
  author={Zheng, Zilong and Liu, Yang and Qin, Shuhan and Lin, Ziyong and Li, Jiaqi and Jia, Zixia and Zhu, Song-Chun},
  institution={BIGAI},
  year={2026}
}