Niels Mündler (@ ICML)
@nielstron
Followers
597
Following
3K
Media
191
Statuses
1K
CS PhD @eth. Language Models, Code, Formal verification. Compiling Python to FP @OpShinDev. Ex-Founder.
Switzerland
Joined November 2012
And this is why benchmark scores keep increasing but my GitHub issues still require manual labor
0
0
1
RT @mbalunovic: We've just released the largest open dataset of expert-annotated LLM proofs! Using this dataset, we did bunch of experiment….
0
4
0
RT @ni_jovanovic: There's a lot of work now on LLM watermarking. But can we extend this to transformers trained for autoregressive image ge….
0
54
0
PLDI'25 is now and I am looking forward to present "Type-Constrained Code Generation with Language Models" in 2 hours (4pm Seoul Time). If you can't make it, feel free check out this thread, which summarizes the key points, or chat me up about any questions.
Excited to present my upcoming PLDI paper at the ICML Workshops DL4C and VerifAI!. Type systems are useful at preventing bugs - so why not leverage them for LLMs? Using constrained decoding, we reduce reduce compiler errors of TypeScript code by over 50%! More details in the 🧵.
0
1
4
Yudkowsky scream going from ultrasonic to. .
Introducing The Darwin Gödel Machine: AI that improves itself by rewriting its own code. The Darwin Gödel Machine (DGM) is a self-improving agent that can modify its own code. Inspired by evolution, we maintain an expanding lineage of agent variants,
0
0
1
I am beyond humbled to see my paper there! glad that people like it :).
0
0
2
so apparently non-big-AI-lab papers *can* trend on Hacker News too.
Suspicious. We put out this tweet thread and then we have the paper on the cover of hn? Correlation, or causation. You decide
1
1
6
RT @mark_veroe: If you are at ICLR, come by our posters in the DL4C (Garnet 218-219) and BuildingTrust (Hall 4 #6) workshops. We have now e….
0
3
0
8/ This work was done in collaboration with my co-lead @jingxuan_he and @MogicianTony @dawnsongtweets @koushik77 @mvechev jointly between @the_sri_lab.and @UCBerkeley. We will present it at ICLR 25 workshops DL4C and VerifAI in Singapore! Excited to discuss it there :).
0
0
2
7/ Most importantly, this method is orthogonal to model scaling - it can be applied on top of however-good models there are. We also see that even SOTA GPT-4o, o1 and R1 make type errors in more complex tasks and languages such as Rust.
1
0
1
6/ The work pays off! For Gemma 2 27B, Qwen 2.5 and DeepSeek Coder we see a drop of around 80% in compiler errors.
1
0
1
5/ Most difficult is deciding whether a partial expression *can* be completed to have a specific type. We implement a search over possible completion types and a heuristic to prune the search space effectively.
1
0
1
4/ We introduce prefix automata to enable reasoning over *eventual* type-safety. These automata track all possible ASTs of a partial program as states. They guarantee reachability of an accepting state (i.e., type-safe program) as long as one AST is found.
1
0
1
3/ We propose leveraging the type system to constrain generation. We prevent sampling tokens from the LLM that violate type-safety. This is hard, appropriate constraints have to be built from scratch for every language. We choose TypeScript as a target due to its popularity.
1
0
1
2/ Prior work demonstrated using constrained decoding to guarantee syntactic correctness, or enforce simple constraints such as correct SQL column names. But syntax errors make up only 3.5% of all code errors in our experiments, and only 6% of compiler errors.
1
0
1
1/ For all the details, check out the paper at We provide a detailed explanation of our constraining approach and extensive evaluation of 6 SOTA open-weight Code-LLMs - which all improve through our method!.
1
0
1
Excited to present my upcoming PLDI paper at the ICML Workshops DL4C and VerifAI!. Type systems are useful at preventing bugs - so why not leverage them for LLMs? Using constrained decoding, we reduce reduce compiler errors of TypeScript code by over 50%! More details in the 🧵.
1
2
5