--- license: apache-2.0 library_name: transformers tags: - agent - rsi - recursive-self-improvement - self-evolving - metaevo - meta-knowledge - swe-bench - swe-bench-pro - deepswe - code-agent - kimi pipeline_tag: text-generation --- # Seshat Agent **Seshat** is a **self-evolving Recursive Self-Improvement (RSI)** agent for software engineering, powered by **MetaEvo** — a self-evolving reasoning infrastructure that turns past successful trajectories into a reusable **meta-knowledge base**, then routes task-relevant meta-knowledge back into the model at inference time. Rather than relying solely on static prompts or fixed workflows, Seshat continuously reflects on task outcomes, compresses useful experience into meta-knowledge, and re-enters the loop so capability compounds over time. On **SWE-Bench Pro**, Seshat evolves a **Kimi K3** base agent from **57.7% → 94.94%** resolve rate. On **DeepSWE-v1.1**, Seshat evolves a **GLM-5.3** base agent from **59.3% → 90.3%** resolve rate. ## MetaEvo: Meta-Knowledge for Task-Conditioned Enhancement **MetaEvo** is a self-evolving reasoning infrastructure. It treats improvement as *experience internalization*, not external skill stacking. Its core idea: once meta-knowledge is abstracted from diverse reasoning trajectories, it can be **recombined and reused** on new problems — analogous to how humans transfer principles across domains (e.g., Newton → Coulomb, Darwin → genetic algorithms). ### Meta-knowledge base The meta-knowledge base stores a large set of meta-knowledge vectors extracted and purified from successful reasoning traces. Each unit is not raw text memory, but a compact, reusable reasoning primitive, for example: - problem decomposition - constraint propagation - reverse / goal-directed reasoning - contradiction elimination - self-verification - tool calling patterns - structural analogy transfer - reasoning-order / adaptive strategy switching Offline, MetaEvo continually mines new vectors from successful trajectories and updates the base — so the agent’s experience becomes a growing prior for future tasks. ### Task-conditioned capability enhancement At inference time, MetaEvo does **not** dump the whole knowledge base into the prompt. Instead it builds a **customized cognitive path** for the current task: 1. **Problem in** — receive the current instance (e.g., a SWE-Bench Pro / DeepSWE bug). 2. **Meta-knowledge router** — match / select / compose the most relevant meta-knowledge units. 3. **Cognitive path** — arrange selected units into a structured reasoning plan. 4. **Inject & reason** — inject the path into the neural forward chain (layer-wise knowledge injection) so the base model reasons with task-specific priors. 5. **Meta-critic** — monitor intermediate states; continue, re-route, or re-plan when needed. 6. **Evolve** — successful traces feed back into the meta-knowledge base for the next round. In short: **current task → retrieve & compose meta-knowledge → enhance the base model’s reasoning on that task → deposit new meta-knowledge**. Seshat is the agent that runs this RSI loop on software-engineering workloads (and transfers the same stack to coding / logic / science benches). ## Key Result | Setting | Base model | Benchmark | Score | |---|---|---|---| | Zero-shot baseline | Kimi K3 | SWE-Bench Pro | **57.7%** | | After RSI evolution (Seshat) | Kimi K3 + Seshat | SWE-Bench Pro | **94.94%** (694 / 731) | | Zero-shot baseline | GLM-5.3 | DeepSWE-v1.1 | **59.3%** (67 / 113) | | After RSI evolution (Seshat) | GLM-5.3 + Seshat | DeepSWE-v1.1 | **90.3%** (102 / 113) | Absolute gain (SWE-Bench Pro): **+37.24 percentage points**. Absolute gain (DeepSWE-v1.1): **+31.0 percentage points**. Evaluation evidence (patches, per-instance results, aggregated scores): - SWE-Bench Pro: [PCL-RSI-lab/SWE-Bench-Pro-Evaluation-Evidence](https://huggingface.co/datasets/PCL-RSI-lab/SWE-Bench-Pro-Evaluation-Evidence) - DeepSWE-v1.1: [PCL-RSI-lab/DeepSWE-v1.1-Evaluation-Evidence](https://huggingface.co/datasets/PCL-RSI-lab/DeepSWE-v1.1-Evaluation-Evidence) ## Evolution Curve The figure below shows Seshat’s RSI trajectory on SWE-Bench Pro: starting from the Kimi K3 zero-shot baseline and climbing through successive self-improvement rounds, eventually surpassing strong frontier agents on the same axis. ![image](https://cdn-uploads.huggingface.co/production/uploads/642d1708c5f19fe0da06b4b7/ERQmh_hqVkZLK0gAzqdD0.png) *Seshat recursive self-improvement trajectory (Kimi K3 base). The curve rises from the zero-shot baseline toward the final evolved performance; dashed lines mark reference model scores.* ## Multi-Benchmark Performance Beyond SWE-Bench Pro, Seshat (shown as **MetaEvo** in the figure) transfers RSI gains across coding, reasoning, science, and IP-matching benchmarks — typically ranking **#1** among the compared systems. ![benchmarks](https://cdn-uploads.huggingface.co/production/uploads/642d1708c5f19fe0da06b4b7/JJ-WV3STxuOwyaelnTOti.png) ### Seshat scores (from the figure) | Benchmark | Seshat | Notes | |---|---:|---| | **SWE-Bench Pro** | **95.1** | Evolved from Kimi K3 (57.0); official submit evidence: **94.94%** (694/731) | | **DeepSWE-v1.1** | **90.3** | Evolved from GLM-5.3 zero-shot **59.3%** (67/113) → **90.3%** (102/113); evidence dataset linked above | | **CyberGym** | **80.3** | MetaEvo on DeepSeek V4-Pro 4-Bit (base 41.2 → 80.3) | | **HumanEval** | **100.0** | #1 | | **StrategyQA** | **82.3** | #1 | | **LogicBench** | **86.4** | #1 | | **LiveCodeBench** | **49.6** | #1 | | **ScienceQA** | **98.0** | #1 | | **SciCode** | **21.2** | #1 | | **PatentMatch** | **76.6** | #1 | ### Selected comparisons | Benchmark | Seshat | Strongest listed baseline | Gap | |---|---:|---:|---:| | SWE-Bench Pro | 95.1 | Claude Opus 5.5 (89.9) | +5.2 | | CyberGym | 80.3 | Claude Mythos (83.1) / GPT-5.5 (81.8) | −2.8 / −1.5 | | HumanEval | 100.0 | Claude Sonnet 5 (98.8) | +1.2 | | StrategyQA | 82.3 | Hy3 (70.1) | +12.2 | | LogicBench | 86.4 | Gemma 3 27B (84.0) | +2.4 | | LiveCodeBench | 49.6 | Doubao Seed 2.0 (41.6) | +8.0 | | ScienceQA | 98.0 | Doubao Seed 2.0 (97.3) | +0.7 | | SciCode | 21.2 | Claude Sonnet 5 (18.8) | +2.4 | | PatentMatch | 76.6 | Qwen3.5 27B (74.1) | +2.5 | ## Intended Use - Research on self-evolving / RSI agents for coding and reasoning - Strong SWE-Bench / SWE-Bench Pro / DeepSWE style repository repair agents - Studying how experience distillation transfers across coding, logic, science, and related tasks ## Citation If you use Seshat or the evaluation evidence, please cite this model repository and the accompanying datasets: ```bibtex @misc{seshat-agent, title = {Seshat: A Self-Evolving RSI Agent for Software Engineering}, author = {PCL-RSI Lab}, year = {2026}, howpublished = {\url{https://huggingface.co/PCL-RSI-lab/Seshat-agent}} } ``` ## Links - SWE-Bench Pro evaluation evidence: [PCL-RSI-lab/SWE-Bench-Pro-Evaluation-Evidence](https://huggingface.co/datasets/PCL-RSI-lab/SWE-Bench-Pro-Evaluation-Evidence) - DeepSWE-v1.1 evaluation evidence: [PCL-RSI-lab/DeepSWE-v1.1-Evaluation-Evidence](https://huggingface.co/datasets/PCL-RSI-lab/DeepSWE-v1.1-Evaluation-Evidence) - Organization: [PCL-RSI-lab](https://huggingface.co/PCL-RSI-lab) ## Contact [seshat.ai2026@gmail.com](mailto:seshat.ai2026@gmail.com)