Ornith-1.5 is a foundation model series that implements a fully integrated, end-to-end self-improvement loop, advancing the self-scaffolding framework from Ornith-1.0. The architecture encompasses three distinct scales—397B MoE, 35B MoE, and a 9B dense model—designed to enhance reasoning, agentic, and coding capabilities. By autonomously proposing tasks, designing task-specific scaffolds, and executing solution rollouts, the model facilitates continuous learning through reinforcement learning.

Core Methodology: The Self-Improvement Loop

The Ornith-1.5 training cycle operates by optimizing the joint pipeline of task generation, harness construction, and solution generation. Unlike static datasets, the system employs a reward-driven approach to expand its curriculum based on capability gaps.

#### Task Reward Function
The reward for a task is defined to ensure it is actionable and effectively pushes the model’s learning frontier. For a question $q$, scaffold $s$, and a set of rollouts $\{\tau_i\}_{i=1}^{N}$, the task reward is:
$$R_{\text{task}} = V(q,s) \times D(q,s,\{\tau_i\}_{i=1}^{N}) \times N(q)$$

• Validity ($V$): A hard gate ensuring the scaffold is verifiable and functional. If $V(q,s)=0$, the task is discarded.
• Frontier Difficulty ($D$): Rewards tasks where the empirical success rate $p$ is close to a target threshold $p^* = 0.2$, defined as:

$$D(q,s,\{\tau_i\}) = \exp\left(-\frac{(p-p^*)^2}{2\sigma^2}\right)$$

• Novelty ($N$): Promotes diversity by calculating the cosine similarity against a buffer $\mathcal{B}$ of previously encountered tasks:

$$N(q) = 1 - \max_{q_j \in \mathcal{B}} \operatorname{sim}(q,q_j)$$

#### Harness and Rollout Optimization
The system evaluates harnesses via $R_{\text{harness}} = C(q,h) \times F(h,\{\tau_i\}) \times H(h)$, focusing on task alignment ($C$), reward fidelity ($F$), and hack resistance ($H$). Solutions are scored based on task success $R_{\text{rollout}}(\tau_i) = h(q,\tau_i)$. The entire three-stage process is optimized using Group Relative Policy Optimization (GRPO).

Performance and Benchmarking

Ornith-1.5 demonstrates state-of-the-art performance in its size class. The flagship 397B model achieves a score of 86.1 on Terminal-Bench 2.1 and 56.0 on DeepSWE, rivaling Claude Opus 4.8. The 35B model notably outperforms dense models like Gemma 4-31B and Meta’s Muse Glimmer-30B on complex agentic coding tasks (e.g., 79.0 vs 52.0 on SWE-Bench Verified). Furthermore, the 9B variant offers strong edge-deployable intelligence, outperforming significantly larger models on key agentic benchmarks. The model consistently maintains performance across coding, reasoning, and real-world agentic environments while utilizing robust anti-hacking protocols, such as removing repository Git history and disabling network access during evaluation.