Bangalore Paper Club - Edition #1
Researchers and builders rarely get into the weeds together. We're fixing that.
Every few weeks, we bring together a small group of researchers, engineers, and builders to go deep on a paper, not just discuss it, but actually work through the math side-by-side.
Expect whiteboard derivations, someone walking through a proof, and the kind of arguments that only happen when everyone in the room has actually read the paper.
This edition, we're covering:
1. The Joint Effect of Quantization and Sampling Temperature on LLM Safety Alignment - A Factorial Analysis (Presented by Hari Prasad, AI Researcher): A large factorial study, proving that quantization barely dents LLM safety while sampling temperature is the real driver of refusal instability, and exposing that models scoring 0% harmful on AdvBench can blow past 80% on harder benchmarks.
2. LLM Inference at the Edge: Mobile, NPU, and GPU Performance Efficiency Trade-offs Under Sustained Load (Presented by Ritam Pal, PhD): The first independent benchmark of LLM inference on a Hailo edge NPU, showing it runs a 1.5B model at under 2 watts with near zero variance and laptop GPU level energy efficiency, while flagship phones shed up to 41% of their throughput to thermal throttling.
3. DeepRetro: Discovers Retrosynthetic Pathways Through Iterative Large Language Model Reasoning (Nature Scientific Reports) (Presented by Rishikesh Panda, AI Researcher): DeepRetro turns a large language model into a retrosynthesis engine that reasons backwards under strict chemical validation, discovering two never before published synthesis routes for complex natural products and winning a $100K chemistry innovation prize.
4. Fast Inference from Transformers via Speculative Decoding (Presented by Ritam Pal, PhD): A tiny draft model races ahead guessing tokens while the big model verifies them in one parallel pass, provably sampling from the exact same distribution as the target, no quality traded away. In practice, 2–3× faster generation with the losslessness guaranteed by construction.
Format
Each paper presented by an attendee (15–20 min), followed by open discussion.
We go line-by-line on the parts that matter. Derivations, assumptions, what breaks in practice.
Small group, capped attendance, so everyone gets to push back and ask the dumb-but-important question.
Off the record. Come to think out loud, not to perform.
Who should come
Researchers, ML engineers, grad students, and builders who want to engage with the actual content of a paper, not just the abstract. Read beforehand if you can to get a lot more out of the room.
What to bring
A laptop or notebook, the paper (links shared ahead of time), and a willingness to be wrong in front of people.
Hosted by The Research Room, in collaboration with Conscious Engines, Bangalore.