Naman Chawla
The Distant Crush Neuron: Single-Neuron Steering of Social Warmth in Qwen3.5-4B
Pushing one MLP neuron (layer 9, neuron 7515) at inference makes a chat character cooler or warmer on demand; blind distance 2.8 vs 1.1 unsteered vs 1.2-2.0 for random neurons; no weight changes · 2026
LG TV Remote for Mac
Menu-bar remote for LG webOS TVs over the local network: buttons, a trackpad pointer, playback and power-on from standby, each confirmed on a real TV · 2026
CAPL: Context-Aware Preference Rules for Steering a Frozen LLM
A small model writes context-scoped preference rules that steer a frozen 9B generator; 71.25% style-rule satisfaction vs 58.33% for a static profile · 2026
Agent-Assisted Reconstruction of Quake II Gameplay Code
Coding agents rewrite decompiled Quake II gameplay functions in C; 99.0% pass a differential harness and the rebuilt module is playable · 2026
Small Models as Agents: SFT and GRPO on a Browser Task
SFT then online GRPO in a live browser takes Qwen3.5-0.8B from 16% to 84% exact-match success on MiniWoB++ click-checkboxes · 2026
Single-Bit Quantization of Transformers/BERT
Benchmarked eight single-bit/ternary quantization strategies inspired by BitNet b1.58; 85.7% accuracy/F1 with STE-quantized BERT on SST-2 · 2025
T5 Knowledge Distillation for Summarization
Distilled a fine-tuned T5-Base summarizer into T5-Small, retaining ~90% of teacher ROUGE and 84.7% of teacher BERTScore · 2025
Mixture-of-Experts Prompt Router
Semantic MoE routing that dispatches prompts to specialized expert models via sentence-embedding cosine similarity · 2025