Akshit Kumar
Hi, I work at Proximal on post-training data, including reinforcement-learning environments and long-context evaluation of coding models. I hold a dual degree in Computer Science and Computational Linguistics from IIIT Hyderabad. I’m broadly interested in how meaning is built in minds and models.
Notes
- William: a tiny poetry model in the browser
- Configuring Pyright with uv
- Interactive Shell Inside a Slurm Job
- Custom uv Environments and Caches with direnv
- Establishing Communication
- --index-strategy unsafe-best-match
- biaffine dependency parsing
- Fixing the Python REPL in Emacs
- std::multiset
- Ada Cluster Reference
- Static Types and Legacy Python
- Fetching Recent Last.fm Tracks
- Exporting Org Notes to Jekyll
- Asahi Linux Setup
Projects
- Reverse Engineering a Stateful Reasoning circuit
Studied Gemma-2-2B on a controlled role-gated retrieval task where a prepositional gate (to or from) selects which of two entities is correct.
- Do Multilingual Transformers Encode Paninian Grammatical Relations?
Investigated how multilingual transformers (XLM-RoBERTa, mBERT, IndicBERT) internally encode grammatical information from the Paninian framework across seven Indian languages. Used diagnostic probing to analyze layer-wise encoding patterns, finding syntactic dependencies peak in middle-upper layers (6-9) while lexical features emerge earlier.
- Adversarial Robustness in Vision-Language Models
Investigated frequency-domain adversarial attacks on Llama Guard 3's multimodal safety filters. Combined vision adversarial attacks with language model security testing.
- Grounding Small Language Models on Structured Data
Explored multi-hop QA using teacher-student distillation for reasoning chain transfer. Built a graph traversal RAG pipeline that achieved 15% accuracy boost by teaching smaller models logical paths.
- A Functional Tensor DSL
Experimented with Racket-based tensor DSL, reimagining ML ops through Lisp's compositional philosophy. Built computation graph auto-differentiation from scratch using functional paradigms.
- A Compiler for a Strictly Typed Language
Nanopass compiler in Racket with graph-coloring register allocation. Implemented tuple support and garbage collection using interference graphs for optimal register usage.