Berkeley, CA/BAIR·Databricks, prev.

Pratyay Pandey

I study what models represent, and how to run them fast.FIG.01

At BAIR, I work with Trevor Darrell and Jacob Steinhardt on mechanistic interpretability: steering vectors and data attribution. Last summer I did LLM inference research at Databricks, on hybrid speculative decoding in SGLang and on model serving.

I read policy and law on the side.

FIG. 01Projection onto a concept direction. Points inside the band are selected; ticks drop to their scalar position along d. The geometry here is illustrative.
Now
2026.08
open questions
  1. 01representation geometrywhat does a concept look like locally?
  2. 02steering vectors as attributioncan a direction point back at the data that made it?
  3. 03acceptance as a probecan draft model acceptance lengths be used to quantify larger model capabilities?
  4. 04steering a genome modelwhat would a tissue-specificity vector steer?
05
2026 · BAIR
Trevor Darrell
Jacob Steinhardt
steering vectors
data attribution
representation geometry

Where does a concept live?

Add a direction to a model's activations and its behavior changes. What I want to know is what that direction implies about everything upstream of it: the shape of the neighborhood it sits in, and which training examples produced the behavior in the first place.

This is attribution done with geometry: rank the training examples by how well they align with the direction, then ask how far up that ranking the responsible ones sit. The work is ongoing at BAIR.

training examples, rankedcos sim →
FIG. 02  Alignment of examples with the direction in FIG. 01. Blue marks the head of an illustrative ranking, where attribution would look first.
06
2026 · Databricks, prev.
LLM inference research
SGLang
model serving
speculative decoding

Hybrid speculative decoding.

At Databricks I worked on hybrid speculative decoding in their internal SGLang, and on serving GLM 5.2 at 381 tok/s within 24 hours of release.

Speculative decoding puts one drafting engine ahead of the large model, and each engine has a range it is good in. N-gram lookup is nearly free and wins whenever the continuation already sits somewhere in the context. EAGLE generalizes where there is nothing to copy. DFlash covers the cases those two handle badly.

The hybrid version keeps them all loaded and chooses per round. Draft with the cheapest engine that is still being accepted, watch its acceptance rate, and hand the slot over once that rate drops. Verification costs one forward pass whichever engine drafted, so the acceptance rate is what decides how many tokens that pass emits.

route → draft → verify → emit—

prefix  …the model can ▮

enginet+1t+2t+3t+4n-gramcontext lookup···EAGLEdraft head···DFlashreserve···emitted····
 
FIG. 03  Two rounds of routed speculation, drawn schematically. One engine drafts per round, and the large model verifies the whole proposal in a single forward pass. Solid blue is a draft the model agreed with. Dashed blue is the token the pass emits on its own, so even a fully rejected round advances generation by one. Each engine keeps the acceptance it earned, and the routing slot goes to whichever is worth its draft cost.
Elsewhere
07 — 09
07Sky Computing Labsystems research with Matei Zaharia
08CS162two semesters teaching operating systems, with Ion Stoica and Matei Zaharia
09Eregiona dynamic fine-tuning system that cut GPU usage 60%, supported by the 1517 Fund's Medici Grant
Notes
10
[01]no entries yet
About
11

I finished my Berkeley EECS degree in a year and a half, mostly so I could spend all my time on research. Along the way, I gained a passion for policy and international law, TA'd for CS 162 with Ion Stoica and Matei Zaharia, and met incredible people!

Outside of work, I write poetry and short essays, am working on my own coffee brand, and enjoy watching movies.