- Joined
- Jan 2026
- How they joined
- Directly
- Before Cognition
- MIT Computer Science and Artificial Intelligence Laboratory (CSAIL)Undergraduate Researcher
- Now
- Thinking Machines Lab
“I helped train it during my internship there. The team is very strong and this is just the beginning!”
- Joined Thinking Machines Lab in June 2026 after graduating from MIT. Verbatim announcement (2026-07-17): "Excited to share that I recently joined @thinkymachines after graduating from @MIT. I have an inkling this will be the next great frontier lab, and a different kind of one. Proud to be working alongside such a fun and talented team."Jul 17, 2026Excited to share that I recently joined @thinkymachines after graduating from @MIT. I have an inkling this will be the next great frontier lab, and a different kind of one. Proud to be working alongside such a fun and talented team.
- Co-author of 'Introducing SWE 1.6: Improving Model UX' (2026-04-07); SWE-1.6 was post-trained from scratch to jointly optimize intelligence and 'model UX', runs up to 950 tok/s, and shipped generally available in Windsurf.Apr 7, 2026
- Listed as a co-author of Cognition's 'An Early Preview of SWE-1.6 and Research Update' (2026-03-01), alongside Carlo Baronio, Ben Pan, Sam Lee, Eric Lu, Steven Cao, Rohan Choudhury, Ray Wang, Gary Chang and Silas Alberti; the post describes a refined RL recipe with two orders of magnitude more compute and SWE-1.6 scoring 11% higher than SWE-1.5 on SWE-Bench Pro.Mar 1, 2026
- Verbatim X post about Cognition (2026-03-01): "New coding model from @cognition! I helped train it during my internship there. The team is very strong and this is just the beginning!"Mar 1, 2026New coding model from @cognition! I helped train it during my internship there. The team is very strong and this is just the beginning!
- Lead author of 'Fast KV Compaction via Attention Matching' (ICML 2026) with Xinghong Fu, Han Guo and Yoon Kim; per his X post it 'achieves 50x compaction in seconds with little performance loss', with code at github.com/adamzweiger/compaction.Feb 19, 2026We introduce a new approach for fast and high-quality context compaction in latent space. Attention Matching (AM) achieves 50× compaction in seconds with little performance loss, substantially outperforming summarization and other baselines.
- Spent Jan-Apr 2026 as a Member of Technical Staff at Cognition working on 'RL for coding models' (LinkedIn description); he describes it as an internship, so it overlapped his final MIT semester.Jan 2026
- Co-lead author (with Jyothish Pari) of 'Self-Adapting Language Models' (SEAL, NeurIPS 2025) with Han Guo, Ekin Akyurek, Yoon Kim and Pulkit Agrawal; SEAL has an LLM write its own finetuning data ('self-edits') and meta-learns that ability with RL.Nov 12, 2025
- Undergraduate researcher at MIT CSAIL (Nov 2024 - Jan 2026), advised by Prof. Yoon Kim, working on data-efficient adaptation and long-context methods for language models.Nov 2024
- Verbatim X post after leaving (2026-07-08): "@cognition congrats @ybenpan @carlobaronio and team!" - congratulating SWE-1.6 co-authors Ben Pan and Carlo Baronio.Jul 8, 2026@cognition congrats @ybenpan @carlobaronio and team!
- Current X bio: 'Working on learning @thinkymachines | prev: @cognition @MIT_CSAIL'.2026
- Cognition signed a definitive agreement to acquire Windsurf, the agentic IDE, in July 2025 after Windsurf's CEO and others left for Google; Adam joined Cognition about six months later and worked on SWE-series models that ship inside Windsurf. He was not a Windsurf employee himself.Jul 2025
- Co-first author of 'The Surprising Effectiveness of Test-Time Training for Few-Shot Learning' (ICML 2025) with Ekin Akyurek, Mehul Damani, Linlu Qiu, Han Guo, Jyothish Pari, Yoon Kim and Jacob Andreas; he maintains the companion repo adamzweiger/Fewshot-TTT alongside ekinakyurek/marc.2025
RL for coding models
https://scholar.google.com/citations?user=MRk2E8AAAAAJ
Excited to share that I recently joined @thinkymachines after graduating from @MIT. I have an inkling this will be the next great frontier lab, and a different kind of one. Proud to be working alongside such a fun and talented team.
@cognition congrats @ybenpan @carlobaronio and team!
New coding model from @cognition! I helped train it during my internship there. The team is very strong and this is just the beginning!
The biggest reduction in KV cache memory comes not from quantization or MLA, but from latent compaction, along the sequence dimension. More strong results coming soon with Attention Matching.
Fun fact: Back in 2014, Demis had a red line condition for any potential acquisition of DeepMind: "no technology coming out of DeepMind will be used for military or intelligence purposes." Google accepting this more eagerly was part of why Demis chose them over Facebook. This red line is even broader than Dario's (no mass surveillance or fully autonomous weapons), though it was quietly removed by Google 1 year ago.
We introduce a new approach for fast and high-quality context compaction in latent space. Attention Matching (AM) achieves 50× compaction in seconds with little performance loss, substantially outperforming summarization and other baselines.
Nice work on GAN-style training with a generator and discriminator, both trained with RL. This might be the path to improvement in domains without good verifiers like creative writing.
Presenting Self-Adapting Language Models on Wednesday at NeurIPS. We equip an LLM with the ability to write training data for itself in response to new inputs. We then meta-learn this ability with RL. Stop by to chat! 11-2 pm, #3415, with @jyo_pari @HanGuo97 @akyurekekin https://t.co/i0KQ7Rw1SL
Interesting work showing concretely why on-policy RL forgets less. It's not quite because of "sparse updates" -- only that RL maintains a smaller KL to the base model.
this is a really fun puzzle :D many more optimizations to come! https://t.co/5V8gb5uRmz
Here are all the architecture tricks used by gpt-oss: - Attention sinks - for each attention head, have a learned scalar such that softmax(qk) becomes softmax over [a_1, a_2, ..., a_T, sink]. Tokens don't have to attend to anything if all the attention scores are low! - Alternate between 128-sliding window layers and dense layers. This is the same as GPT-3. - top-k MoE, with no shared expert and k=4. Standard load balancing loss. - Really nonstandard swiglu - Uses σ(αx) where α=1.702 which makes the silu approximate a gelu. Clips gate activations to (-inf, 7], and up activations to [-7 and
Models that know what they know are way more useful. So, jointly reward correctness and calibration!
