Yuvanesh S

SpecQuant is on arXiv now!! arxiv.org/abs/2609.21704 uses multiple quantized variants of the same parent model for speculative decoding, with complexity-based routing for faster local inference (building on the routing idea from our FrugalSOT work). did these experiments around a year ago with @scriptosis @harish20205, so some of the setup may be a bit dated now :)

almost a year ago, me, @scriptosis and @harish20205 were sitting on a train reading the EAGLE-3 paper and discussing this idea (and we’re now going to present our first paper in a conference). a lot of EAGLE-3's performance still depends on params you pick yourself, so we were thinking, why not learn to pick those params based on the actual hardware/load instead? finally got around to actually trying it out :) wrote a small blog on what came out of it: pradheep.dev/blogs/energy-ba…

following up on this, the experiment i've been trying is to distill the knowledge from the strongest q0 teachers based on our compute budget K. the intuition i had for this was MOPD. they distill different skill/domain based teachers into a single model. here, instead of different skills, we have different models that have seen the same data but ended up understanding it in slightly different ways. tried this with Qwen2.5-1.5B, trained only on GSM8K, and these are the results on MATH-500. this is from a top-2 mixture out of the 4 cyclic snapshots

one of our papers, 'alphadesign: a hybrid reinforcement learning and genetic algorithm approach for f1 front wing optimization compliant with fia 2026 regulations', got accepted at flins-iske 2025 couldn't afford the registration so we won't be attending. trying to put it up on arxiv instead -- if anyone can help with an endorsement for cs.LG or cs.NE, would really appreciate it huge credit to @scriptosis, @harish20205

Just made understanding any codebase way simpler. DGAT is live: uv pip install dgat dgat config init dgat scan /path/to/your/project One command and you get: - file_tree.json - dep_graph.json - dgat_blueprint.md (LLM-synthesized architectural overview) Point it at any repo → instant annotated dependency graph + blueprint. Trying with different harnesses like opencode, claude code. Benchmarks coming soon!!

One of our papers, 'SpecQuant: Speculative Decoding with Multi-Parent Quantization for Adaptive LLM Inference', just won the Best Paper Award at a conference! Huge credit goes to @harish20205 and @scriptosis !" project website: specquant.vercel.app/

Just recreated the ImageNet RL experiment from the MAXRL paper myself - ran GRPO, classic REINFORCE, and MAXRL head-to-head on the same setup. Final metrics came out like this: - GRPO: best loss -0.0010 | best acc 0.0386 | final acc 0.0386 - REINFORCE: -0.0016 | 0.0361 | 0.0361 - MAXRL: -0.0017 | 0.0407 | 0.0407 MAXRL edges out on accuracy which tracks with the paper's claims—going direct for likelihood maximization instead of proxy reward feels cleaner. Tracked validation acc/loss over epochs + per-step loss curves too. The per-step one is noisy af as expected with Monte-Carlo sampling but you can see the trends. Attaching the terminal output, val acc plot, loss/epoch, appendix snippet explaining the advantages, and the raw loss/step graph. Feels good to close the loop on this one after geeking out over the paper last week 🤯 Blog coming soon!