Yves Raimond

A new paper from SJTU and GAIR (daVinci-Env) makes a compelling case that we've been looking at the wrong variable in SWE agent training. The premise is refreshingly honest: the open community doesn't lack clever RL algorithms or agent architectures. What it lacks is a large-scale supply of executable, verifiable training environments. So they built one. 45,320 Docker environments across 12.8k Python repositories, synthesized from 572K GitHub PRs through a multi-agent builder pipeline. The environments come with Dockerfiles, evaluation scripts, and distributed construction infra - all open-sourced. The engineering decisions matter more than the headline numbers: -> Builder decomposed into retrieval / setup / evaluation / analysis roles (not one monolithic agent) -> Dual-run validation protocol: test-only must fail, test-with-fix must pass -> Structured failure routing that becomes free process supervision for training -> Success-frequency curriculum: keep instances where the model succeeds 1-2 out of 4 attempts The most convincing result is the environment-source ablation: holding model and scaffold constant, swapping only the training data source, OpenSWE beats SWE-rebench in all four settings - up to +12.2 points. That's direct evidence that environment quality is a first-order training variable. Where the paper is weaker: the "difficulty-aware curation" story lacks clean ablation, the 32B headline slightly oversells relative to the strongest RL baseline (+1.0, not +4.6), and whether the curated trajectories are fully released remains unclear. The real takeaway for anyone building post-training pipelines: verifiable feedback loops and environment quality may matter more than the choice of optimization algorithm. This paper has quietly assembled every ingredient needed for an RL-for-code environment engine - real repos, executable sandboxes, verifiable outcomes, structured error attribution. The next step is online policy improvement on top of that substrate. https://lnkd.in/e2cFM-zu

5

This is a great experiment by the Anthropic team, showcasing the "true autonomous" capabilities of LLMs. By "true autonomous," I mean that Anthropic (I believe) purposefully did not bombard the LLM with numerous tools in areas where LLMs are weak. For example, LLMs are not good in math, and performance would have been much better if they hd provided a tool like "calculate_profitability." This might have prevented the $100 opportunity miss. In AI agent development, tools bridge the gap where LLMs are not "deterministic." This is currently the MOAT in the GenAI app business: the better your tools, the more deterministic results you can provide. Only till LLMs can write their own tools on the fly and execute them where they know they are weak and using deterministic softwares (tools) to fill these gaps. https://lnkd.in/gMk3VRJv

10

1 Comment

To FTP or to sFTP? Been using these technologies for 30+ years, and somehow I find myself still wondering when to use one or the other. The folks at #kiteworks have a nice article: https://gag.gl/oi1wa2 SFTP (Secure File Transfer Protocol) offers a highly secure method for transferring files by using encryption, authentication, and data integrity checks, making it a significant improvement over the older, unencrypted FTP protocol. However, it comes with drawbacks such as increased complexity, slower performance, and limited native support for compliance features like audit logging or encryption at rest. Tools like Kiteworks enhance SFTP by adding security, governance, and compliance capabilities, making it more suitable for regulated environments.

4

The shift from metric-based to narrative CVs is fundamentally changing how research is funded. But asking scientists to suddenly become storytellers introduces a massive friction point. For funders like Research Ireland and UK Research and Innovation(UKRI), the days of relying solely on an h-index are fading. They want to see real-world impact and collaboration. But translating years of highly technical research into a compelling narrative format takes a significant amount of time. To help solve this, I partnered with a brilliant engineer to independently build and launch DORA CV Builder (www.doracv.com). It’s a free, AI-powered tool that uses the STARS framework to guide researchers in extracting the actual story and impact behind their work. When defining the product requirements, we recognized a major adoption barrier: intellectual property and data privacy. Researchers cannot risk their pre-published data floating around the cloud. So, we made a strict product decision regarding data handling: your content is processed securely in real-time, and absolutely no data is stored or retained on our servers. If you are a researcher navigating these new funding requirements, give it a try. It's completely free, and I’d love to hear your feedback on the UX and how it fits into your workflow. #NarrativeCV #AcademicChatter #ResearchAssessment #ProductManagement #BuildInPublic

38

A new paper identifies typicality bias in human preference data as a primary driver of mode collapse in aligned language models. The research proposes Verbalized Sampling as a training-free solution. The Problem: Post-training alignment processes like RLHF reduce model output diversity, causing models to favor narrow sets of responses. Previous work attributed this to algorithmic limitations, but this research traces it to a data-level issue: human annotators systematically prefer familiar text patterns, a phenomenon rooted in cognitive psychology. The Solution: Verbalized Sampling prompts models to generate probability distributions over multiple responses rather than single outputs. For example: "Generate 5 jokes about coffee and their corresponding probabilities." Results: Testing across GPT-4, Claude, Gemini, and open-source models showed: - Creative writing tasks: 1.6-2.1× increase in diversity compared to direct prompting - Improved performance in dialogue simulation, open-ended QA, and synthetic data generation - No reduction in factual accuracy or safety Stronger models showed greater benefit from the technique The method requires no model retraining and works at inference time. The research provides a data-centric framework for understanding alignment trade-offs and offers a practical approach for applications requiring diverse outputs. https://lnkd.in/g3d-_nDM

3

1 Comment