Did some SFT on Qwen3.6-35B-A3B this evening, just to get it to respond better to Jev-y prompts via github.com/ekzhang/openjev-s… Cost: $5, easy 10 min training run on Tinker. +8% on GPQA diamond and +12% on MMLU-Pro. Also evaled @jaredpalmer's Kev here for comparison

Sharing my first of hopefully many research blog posts! jasperlu.com/blog/training-s… This one is the kind of educational blog post I wish I'd had when I started with RL for LLMs. I tried to make it as open as possible. Every rollout is browsable, the code is open source, and I walk through my entire thought process, from learning rate sweeps to reward shaping.

We fine-tuned @thinkymachines' Inkling-Small to fix LaTeX compile errors inside the editor, in under a second. On real errors it matches Claude Fable 5.1, six times faster and at about 1/40 of the cost. When we interviewed mathematicians, one of the most common annoyances we heard about in their day to day work is battling with LaTeX compilation errors. The most obvious solution is to ask a chatbot to fix errors. That works decently well, except switching to a different window breaks the writing flow, and copy-pasting the snippet often loses relevant context, especially when multiple files are involved. So we scraped 272,000 questions from TeX.StackExchange and kept the 3,978 whose snippet still fails today, whose accepted answer compiles, and whose difference replays exactly as a patch. The largest dataset previously released had 88. We tried tweaking the reward function in many ways. The first one was only a compilation check, which led the model to remove content until the document built. In the end, what worked best was rewarding compilation and a close match with the solution's PDF, and penalizing removed content. We're rolling out the model inside the @sundialmd editor this week. When a compile fails, the fix is applied as a suggestion in under a second and the PDF rebuilt. Write-up: sundial.md/blog/textinguishe… Thanks @tinkerapi and @thinkymachines for the support.

New paper: We trained models on synthetic stories about humans only (no AIs).
 We found the Assistant adopts quirky behaviors from the stories in ordinary chat. Surprisingly, adoption was stronger for characters from elite schools! Why does this happen? 🧵

AI is now improving its own model architectures and chips. The power infrastructure it runs on deserves the same attention. We used RL to teach a Kimi base model to design medium-power transformers. It met 93% of unseen specs, helping to compress multi-week engineering efforts into a few minutes of inference. 1/5

Using Tinker with an autoresearch loop is a really effective way to reproduce post-training papers at predictable costs Today there are dozens of self-distillation methods all claiming improvements over each other, and it’s hard to establish which claims hold up We gave agents a Tinker budget to reproduce self-distillation results across models and training setups. With just a few user prompts, they reproduced SDFT’s continual learning benefits across Qwen3-8B and Qwen3-30B-A3B over multiple seeds, and investigated SFT’s failure modes. Read more below:

The causes vary widely: specific words in the prompt, quirks of the model (a misremembered fact about an actor), or abstract properties (a user's angry tone).

How to post-train a model to personalize it on your code repo? In our latest research in Bespoke Labs, we post-trained a model to improve its performance on a given Github repository. Starting from Inkling base, we use supervised fine-tuning (SFT) with trajectories coming from a strong teacher model, and reinforcement learning (GRPO) on repository-specialized environments that we curated. SFT gave a 52pp improvement in performance on the held-out fontTools evaluation set. Further RL training lifts the total improvement to 57pp compared to the base Inkling model. In addition to the in-distribution evaluation our post-trained Inkling shows good performance on Terminal-Bench 2.1 and SWE-Bench Lite while becoming 40% more token efficient due to post-training. Read our full research blog post here: bespokelabs.ai/blog/personal… Many thanks to Thinking Machines Lab for their credit contribution that helped support this research.

New preprint from @lightningrodai, this time with Philip Tetlock and Ville Satopää! We post-train 5 versions of the same LLM, changing only the scoring rule used as the RL reward. Similar aggregate scores, but very different BIN profiles. A good Brier score alone doesn't tell you if a forecaster can distinguish likely from unlikely events. One might discern well but lose if its probabilities run systematically too high. A less discerning one might score better by hugging the base rate. BIN splits forecast performance into bias, information, and noise. Bias is a systematic shift in the probabilities. Noise is random scatter. Information is real signal about which outcomes are more likely — the part you need a powerful LLM for. Different uses call for different profiles. Reward choice is one lever shaping which forecaster you get. Congrats to co-authors @indiequant @KSkotheim64001 @VSatopaa @PTetlock 🙌 Full paper: arxiv.org/abs/2608.28482

Introducing autoresearch for arXiv papers: replicate and experiment on any arXiv paper using an army of Claude/Codex agents For post-training, your agents can use @tinkerapi to launch concurrent RL runs while you watch your experiment tree grow Try it out now on OpenResearch: openresearch.sh

We use Inkling-Small to turn paper abstracts into quick, useful summaries. open weights × open science 🤝

Recently we are working on adding proactive support to the collaboration layer. While testing with different people, I found audio is a good way to capture moments when people need support. So we added a "Hi Coco" feature and support inkling model through @tinkerapi - it works pretty well! (video has sound!🔊) We open source this part first since several people told me they're exploring ideas around collaboration layers + talk-out-loud. Link in thread.