One-Line Summary

Drug discovery gets more disease-aware while vision research pushes longer, steadier videos and safety-focused benchmarks, with an open-source agent hub streamlining prototyping.

Research Papers

DrugGen-2 pairs disease context with targets to design drug candidates

DrugGen-2 is a generative system that proposes small molecules while explicitly considering which disease they are meant to treat and the target protein’s sequence, rather than optimizing chemistry in a vacuum. Built by fine-tuning a Generative Pre-trained Transformer 2 (GPT-2) and then applying reinforcement learning, it aims to produce compounds that are valid, novel, diverse, and likely to bind strongly. ¹

The training recipe uses supervised fine-tuning followed by Group Relative Policy Optimization (GRPO) — a reinforcement learning step guided by reward functions for chemical validity, novelty, diversity, and predicted binding affinity — with conditioning on both disease ontology and target sequences. This setup turns disease context into a first-class input alongside the protein target. ¹

On five protein targets tied to diabetic nephropathy, DrugGen-2 outperforms baselines DrugGPT and DrugGen: it yields more unique molecules, higher structural similarity to approved drugs, and stronger predicted binding across all targets. Molecular docking highlights candidates with predicted affinities of -9.917, -9.485, and -9.367, surpassing a reference drug (enalapril at -8.283). ¹

By injecting disease-specific signals into generation, DrugGen-2 points toward more realistic de novo design and drug repurposing workflows that reflect how targets behave in different conditions — a step that can reduce blind spots in early screening. Watch for broader evaluations beyond diabetic nephropathy and comparisons across additional disease ontologies. ¹

LongE2V stabilizes long-horizon video from event streams with diffusion priors

LongE2V reconstructs high-quality video from sparse event streams and also handles future prediction and in-between frame interpolation, all by fine-tuning a pre-trained video diffusion prior. In plain terms: it turns the rapid, change-only readings from event sensors into coherent long videos while filling in what happens next and between frames. ²

To keep sequences stable over long durations, the method introduces Autoregressive Unrolling and Adaptive Context Switching; for interpolation consistency it adds Reencoding Alignment with Cross Residual Correction, and for robustness across sensors it uses Event Voxel Density Augmentation. Across real-world benchmarks, the paper reports state-of-the-art results on reconstruction, prediction, and interpolation, with strong temporal coherence and zero-shot generalization. ²

AUTOPILOT-VQA tests dashcam incident reasoning in vision-language systems

AUTOPILOT-VQA is a benchmark that checks whether models can answer grounded questions about driving incidents in dashcam videos — moving beyond object naming toward incident-centric reasoning. It targets Vision-Language Models (VLMs), Large Language Models (LLMs), and Multimodal Large Language Models (MLLMs) used in autonomy stacks. ³

The dataset asks structured questions about real incidents and near-incidents, covering weather, lighting, traffic environment, road layout, surface state, signage, involved entities, whether an accident occurs, impact location, and whether it could have been avoided. Released as part of the AUTOPILOT CVPR (Conference on Computer Vision and Pattern Recognition) 2026 competition, it offers a standardized way to assess reliability across scenarios and nudges models toward temporally grounded, safety-aware reasoning. ³

Canvas360 uses geometry-aware pretraining for in-context panorama generation

Canvas360 is a two-stage framework for panoramic image generation that first learns geometry-aware priors and then fine-tunes for in-context tasks like style transfer, inpainting, outpainting, and editing — aiming to keep global geometry consistent while handling local edits. ⁴

It introduces Canvas360Dataset with 1 million paired panoramic samples and modeling elements such as parallel depth generation, velocity circular padding, and similarity loss regularization. A unified, token-level concatenation interface lets the same model handle multiple panorama tasks, and experiments show improved fidelity and geometric coherence on panorama-specific metrics with competitive or leading quantitative results. ⁴

Open Source & Repos

botpress consolidates tools to build and deploy LLM agents

Botpress is an open-source platform for building chatbots and AI agents powered by providers like OpenAI, with quick starts via its Software Development Kit (SDK) and Command-Line Interface (CLI) packages (@botpress/sdk, @botpress/cli) and a managed Botpress Cloud. It positions itself as a hub to create and deploy assistants backed by Large Language Models (LLMs). ⁵

The GitHub project links to documentation, community channels (Discord, YouTube), and a cloud console, making it approachable for teams prototyping customer support, FAQ, or workflow assistants and then moving to hosted deployment. ⁵

Why It Matters

When generation is tied to the real constraint space — a disease ontology, a physical sensor stream, or incident structure — models tend to produce outputs that better match how decisions are made in practice. DrugGen-2, LongE2V, and AUTOPILOT-VQA all channel domain context to reduce guesswork and highlight failures earlier. ¹

On the tooling side, consolidators like Botpress cut setup time for teams that need to stand up assistants quickly, complementing research advances with faster paths from prototype to pilot. ⁵

What to Try This Week

1. Botpress quickstart: Clone the GitHub repo and use @botpress/cli to scaffold a simple FAQ bot, then test deploy to Botpress Cloud (https://github.com/botpress/botpress).