Agentic Bigbang

Data, evaluation, and post-training for agents that work.

Agentic Bigbang brings together our research on agentic data synthesis, data processing, evaluation, and post-training across two domains:

• SWE: agents for repository-level software engineering, from task organization and expert training to single-policy integration.
• Cowork: agents for knowledge work and productivity tasks, including document-grounded slide generation and task-specific evaluation.

This is a shared research repository. Each work has its own documentation, artifacts, and release scope; reusable tools live alongside the paper packages.

Two Logics-SWE models, one set of three category experts. Our 27B and 35B-A3B students bring category-specific expertise into single-model coding agents through multi-teacher on-policy distillation (MOPD).

Improvements are over the corresponding Qwen3.6 base model.

Paper · 27B model · 35B-A3B model · Training data and environments · Evaluation data · How it works

Public releases are listed newest first. Dates are in UTC; future papers, models, datasets, and tools will be added as they become available.

Paper · Artifacts & reproduction · SWE Labeler · Model card · Figures & supporting data

TL;DR: Develop category specialists through Refresh–Repair–Expand (RRE), then consolidate their behaviors into one deployable agent with multi-teacher on-policy distillation (MOPD).

Repository-level software engineering spans heterogeneous tasks. Under pooled reinforcement learning, gains in one category can coincide with regressions in another: aggregate resolution can hide this category see-saw. Simply splitting the training data by category does not guarantee stronger experts. Our work therefore studies how to develop and integrate category experts, rather than treating category splitting itself as a demonstrated improvement.

1. Organize tasks with SWE Labeler. Evidence-grounded, multi-axis labels organize executable tasks into three repository-domain categories and support category-level training and analysis.
2. Develop experts with RRE. Starting from the same base model, category experts alternate long-horizon Agentic-miniRL with mastery refreshes, supervised repair using their own verified successful trajectories, and task-pool expansion for further RL.
3. Integrate with MOPD. A student generates its own trajectories and receives label-routed supervision from the category experts, with reference-anchored extrapolation. The resulting policy requires neither category routing nor multiple expert models at inference time.

RRE develops the category experts; MOPD distills them into a single student. Integration is on-policy distillation, not arithmetic averaging of weights.

Expert training and policy integration use no external model to supply solution trajectories or action targets. This statement does not exclude the use of a labeling model for task annotation.

SWE Labeler is a reusable annotation tool for both repository-level task instances and agent interaction trajectories. Its evidence-grounded taxonomy describes what a task changes, the repository's engineering domain, and the scale of the work:

• Task Type: 26 L1 families and 119 fine-grained L2 labels.
• Repository Domain: 21 L1 families and 108 fine-grained L2 labels.
• Three scale axes: modification scope, cognitive complexity, and estimated resolution time, each with four levels.

The paper's taxonomy atlas shows the label inventory, not dataset frequencies. Click the image for the original vector PDF.

In RRE–MOPD, a fixed map groups repository-Domain labels into the A/B/C categories used for expert training, teacher routing, and category-level evaluation. Task Type, fine-grained labels, and scale annotations remain available for data profiling and training diagnostics. The labeler can also be used independently of the RRE–MOPD training pipeline.

Code & quick start · Taxonomy definitions · Benchmark annotations

Our Logics-SWE family brings category-aware expertise into single-model coding agents. Both students learn from the same three 27B RRE experts, covering service, data and security; user-facing applications; and systems, tooling and runtimes. MOPD transfers their task-specific strengths through on-policy distillation, so the resulting agents do not need teacher routing at inference.

• Logics-SWE-Qwen3.6-27B is our dense 27B student, initialized from Qwen3.6-27B.
• Logics-SWE-Qwen3.6-35B-A3B is our MoE student, initialized from Qwen3.6-35B-A3B and trained with the same 27B teachers. It extends the approach to a different student architecture. Model card.

The expanded evaluation measures task resolution over three rounds. See the benchmark overview for results across four evaluations and the Pro V2 Hard-51 subset.

The original paper reports 58.04 ± 0.20% for Logics-SWE-Qwen3.6-27B and 52.64 ± 0.28% for Qwen3.6-27B on Pro-618, a 618-task subset of SWE-bench Pro. Values are the mean ± population standard deviation over three evaluation rounds. See the aggregate results for the underlying table.

• Download the model: Logics-SWE-Qwen3.6-27B on Hugging Face, the released single policy trained with RRE–MOPD.
• Use the released instances and environments: Logics-SWE-Env-2.5K on Hugging Face provides 2,553 instance records and matching container references; see the download and execution guide.
• Inspect the experiments: training manifests and evaluation metadata, per-task results, and training and evaluation configurations.
• Reproduce the analysis: validation and plotting instructions, statistical reproduction, and figure index.
• Use the labeling tools: SWE Labeler, taxonomy definitions, and benchmark annotations.

Paper · Research overview & case studies · PPTSynth code · Released models

TL;DR: Logics-PPT develops agents that turn source material into complete, visually coherent presentations. The training recipe combines grounded task and rubric synthesis, guided trajectory distillation, and reinforcement learning with rewards measured on rendered slides.

Presentation generation asks an agent to understand source documents, organize faithful content across a deck, and make each slide readable and well composed. An attractive slide can still omit key facts; a faithful slide can still have colliding elements or unreadable text. Our work treats these as connected training problems, from constructing supervision to checking the final render.

The illustrated pipeline connects three research stages. The released PPTSynth package covers task and rubric synthesis; the SFT and RL model checkpoints are linked below.

1. Ground training tasks in documents. Build source cards, slide-generation instructions, and instance-specific binary rubrics for material completeness and correctness. Audit weak items and refine them against source evidence.
2. Teach the agent through guided distillation. Use explicit content and design guidance, complementary teacher trajectories, and rendering-based filtering to train dense and mixture-of-experts student models.
3. Refine what people see. Render generated HTML slides and use five verifiable rewards—aspect ratio, whitespace, element collision, visual balance, and font readability—to guide reinforcement learning.

The technical report includes representative, unaltered slide outputs from before and after training. These examples illustrate changes in deck structure and slide layout; they are not a substitute for benchmark results.

Training stage	Before	After
SFT: more structured slide content
RL: resolve element collisions

Explore all six before/after comparisons and the method.

Four research checkpoints are available for slide-wise HTML presentation generation. SFT models focus on the full presentation-generation task; RL models further target post-render visual quality. Full generation depends on the agent and rendering setup described in each model card.

PPTSynth is the released code entry point for grounded task and rubric construction. It supports English and Chinese academic papers, English finance/economics reports, and Chinese general-domain documents. See the quick start and data policy.

• Repository artifacts: SWE Labeler and PPTSynth code, the RRE–MOPD research package, and the Logics-PPT overview and figures.
• Models: Logics-SWE-Qwen3.6-27B and Logics-SWE-Qwen3.6-35B-A3B and four Logics-PPT SFT and RL checkpoints on Hugging Face and ModelScope. Each model card gives its license and usage details.
• Data and environments: Logics-SWE-Env-2.5K provides 2,553 instance records with matching environment-image references and a documented execution protocol.
• Papers: One to More, More to One and Building Presentation Intelligence for High-Quality Slide Generation are available on arXiv.
• Not included: complete SWE training data and execution trajectories, or Logics-PPT source documents, generated tasks, full training trajectories, and unpublished training or reward services. The anonymous SWE manifests document training membership but do not contain the underlying task content.

Licensing is component-specific: SWE Labeler and PPTSynth include Apache-2.0 licenses. See the SWE Labeler license scope for its coverage and third-party exceptions.

These code licenses do not automatically license other research artifacts, third-party datasets, or model weights. Consult the applicable component and artifact notices. The released Logics-SWE-Qwen3.6-27B model is licensed under Apache-2.0.