Alex Zhang

Alex Zhang is an MIT researcher studying how the systems around language models change what they can do. With Tim Kraska and Omar Khattab, he coauthored work on Recursive Language Models, a way to let a model use code and smaller model calls to work through material too large to fit in one prompt. He has also coauthored KernelBench, a benchmark for generated GPU code.[1][3][4]

A recursive language model is a harness, or program around a model, rather than a new neural network architecture. The work by Zhang and his coauthors puts a long input into a persistent programming environment. The model can write code to inspect portions, call another model on selected pieces, store intermediate results, and combine them into an answer. For example, a question spanning many documents can be broken into smaller questions whose results are processed in code.[3][1]

The paper evaluates the approach on several long-context tasks, including a multi-document research task, repository understanding and OOLONG-Pairs, which requires aggregating information across pairs of entries. Its authors compare against direct calls and alternative scaffolds. These are results on selected tasks and systems, not a guarantee that recursion produces a correct answer: the program still has to select the right material, solve the subproblems and assemble their results.[3]

In his 2026 Latent Space interview, Zhang connects this design to compositional generalization. He describes a training experiment where the same high-level program can handle a longer problem by changing a length variable, while subagents work on smaller pieces. He also proposes that similar strategies could transfer across different domains; that broader possibility remains a research question, not a result for every kind of task.[1]

Earlier, Zhang worked with the SWE-bench team at Princeton and participated in GPU Mode. He is one of seven authors of the 2025 KernelBench paper, which evaluates whether language models can produce correct, fast kernels across 250 selected PyTorch workloads. The authors found that their tested models still struggled to match the baseline across many cases; the benchmark measures the specified workloads and hardware setting, rather than all GPU programming.[1][4]

Zhang distinguishes the speed of an isolated kernel from the speed of a full model. One operation might run fastest by itself when its output is written to high-bandwidth memory; accepting a slower first operation could leave data in cache for the next one and improve the combined computation. He also describes an AI-assisted leaderboard solution guided by a skilled GPU Mode member that was unusually stable in an end-to-end system. His example illustrates the need to check integration and correctness alongside a narrow speed score.[1]

Zhang says he collaborated with Prime on Prime Agent, an implementation using the Pi harness with IPython as the direct tool; other capabilities load as Python modules or scripts. Its context is also saved on disk, although it retains a conventional conversation and compaction loop. He describes agent-to-agent communication and persistent subagents, and credits a separate continual-harness idea to another researcher, Seth. He says he was not involved in Prime’s model-training work.[1]

The tradeoff is cost and latency. Zhang says compaction can be quicker and cheaper even when an RLM can address a harder task. He treats the idea that more elaborate agent systems will expand model capabilities as a research bet, not a reason to assume that the extra coordination is worthwhile for every request.[1]

Zhang’s personal site identifies him as an MIT CSAIL PhD student and a 2024 Princeton computer science graduate; it names Omar Khattab and Tim Kraska as advisers. He traces his interest in GPU programming to a Princeton talk on FlashAttention, and says he spent a summer at Sakana AI before his PhD. His work spans benchmark design and the question of how a model’s surrounding software changes the tasks it can tackle.[2][1]