|
— |
Introduction |
11:00 AM – 12:00 PM |
— |
| 1 |
Paras Chopra |
Lossfunk Research Previews |
12:00 PM – 12:30 PM |
— |
| 2 |
Kushan Raj |
Beyond Pretraining: Building Robots That Adapt Across Embodiments |
12:30 PM – 12:45 PM |
Current VLAs can reason across embodiments thanks to massive pretraining. This is only going to improve due to the recent scale-up in collecting ego-centric data (video data). But high-quality embodiment data is still the bottleneck to extend capability to new embodiments, and to new tasks in those embodiments.There are also no good benchmarks for cross-embodiment adaptation — everyone tests on their own hardware with their own tasks. Setting up a clean comparison in Robosuite (train on Franka, deploy on UR5e) could be a useful output on its own. |
| 3 |
Chaitanya Sharma |
Less is More? Multi Agent Systems - What, When, Why, Where |
12:45 PM – 1:00 PM |
A counterintuitive look at why stacking AI agents often makes systems dumber and the precise conditions under which cooperation actually pays. |
| 4 |
Tushar Magar |
Babylon: Designing the Future of Live Events |
1:00 PM – 1:15 PM |
After quitting my job at a YC startup to build my own, I spent my time at Lossfunk working on Babylon: a creative technology company starting with laser automation, but moving toward the broader automation of lights, fixtures, and live event systems.This talk is about my builder journey while at the programme, and how Babylon is evolving into a tool for making clubs, concerts, and live events more immersive, intelligent, and unforgettable. |
| 5 |
Aadi Narayana Varma Dantuluri |
Measuring Whether the Scholarly Record Is Ready for AI-Mediated Use |
1:15 PM – 1:30 PM |
AI systems can only reliably discover, summarize, and credit scholarly work when research metadata is complete. Using OpenAlex, this study introduces the NEXUS Score and shows how missing metadata leads to citation errors and hidden contributors that cannot be fixed with more compute or web search.The core argument is that metadata repair is a one-time infrastructure improvement that benefits every future AI system. |
|
— |
Lunch Break |
1:30 PM – 2:30 PM |
— |
| 6 |
Pranay Kundu |
How to benchmaxxxx |
2:30 PM – 2:45 PM |
What are the common techniques used by AI labs to climb benchmarks methodically. |
| 7 |
Harshvardhan Vatsa |
Investigating Normalization in Mamba-3: What BCNorm Computes and What It Costs |
2:45 PM – 3:00 PM |
Mamba models normalise their internal "memory keys" at every layer, for every token.. the slowest single step when generating text. I trained two identical models, one with and one without normalisation and opened them up to see what it really does. |
| 8 |
Pratik Pattanaik |
I think therefore I speak |
3:00 PM – 3:15 PM |
I want to talk about my research on neural signals decoding, paper replication done on the same, some multi architecture ablation experiments, and an exploration into speech neuroprosthesis to study robust real time streaming of thoughts to text to help aid ALS patients. |
| 9 |
Shreya Jain |
Moe inference optimisation on edge device |
3:15 PM – 3:30 PM |
This past 6 weeks I have been looking into learned MoE expert-prefetch predictor to deliver real wall-clock decode speedup on commodity hardware (T4, Mac M1 Pro) |
| 10 |
Viraaj Gupta |
Attention Doesn't Understand Itself |
3:30 PM – 3:45 PM |
We've spent years teaching models what to learn. The next question is whether they can learn how to learn it. I spent 6 weeks at Lossfunk working on that frontier, I think that's where the next generation of architectures lives. |
| 11 |
Kumar Kartikay |
The AlphaZero Playbook: Integrating Tree Search with Deep Learning in LLM Agents |
3:45 PM – 4:00 PM |
Inspired by AlphaZero's breakthroughs in Go and Chess, this talk dives into how we can bridge explicit tree search and deep learning to build LLM agents capable of far more sophisticated, multi-step reasoning. |
|
— |
High Tea Break |
4:00 PM – 4:30 PM |
— |
| 12 |
Kanak Raj |
When LLMs Forget the Latest: Failures in State Tracking in KV streams |
4:30 PM – 4:45 PM |
LLMs consistently retrieve a variable's initial value over its most recent one -a bias seen across 20 models and up to 53 percentage points. A tiny LoRA adapter (0.24% of weights) restores accuracy to 93–100%, and checkpoint analysis shows models internally track updates but suppress them during decoding. |
| 13 |
Shreyas Ramachandran |
The Path From World Models to Autonomy: Open Research Questions |
4:45 PM – 5:00 PM |
In this talk, I will explore the key challenges that remain in achieving true autonomy, examining world models, memory architectures, and their limitations. I will argue that future autonomous systems will require a deeper understanding of causality, creativity, and embodied interaction with the environment.I will conclude by introducing neural computers, a paradigm that unifies I/O, compute, and memory within a single framework. |
| 14 |
Dr Raunaq Pradhan (Online) |
Who's Asking? |
5:00 PM – 5:15 PM (Online) |
Ask an AI the same medical question as a doctor, then as a patient - and the answer can change. Same model, same evidence; only the asker changed. A short investigation into when AI bends to its audience, why it happens most where experts disagree, and what that means when the person asking can't push back |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|