Peter Massih

Background

Publications

SQuID: A Benchmark for Quantitative Spatial Reasoning on Satellite Imagery

P. A. Massih, E. Cosatto, M. R. Min · Preprint, arXiv:2601.13401, 2026 · arXiv →

Three Modalities of Production AI in Financial Services: Retrieval, Generation, and Detection

P. A. Massih · Master's thesis, EPFL, 2026 · DOI →

All publications →

Selected projects

3D-BlockGen: turned text into buildable LEGO models with a two-stage diffusion model (shape, then color) trained on 542,292 voxelized Objaverse samples. It beats single-stage variants on IoU, F1 and color.

LLM Teaching Assistant (EPFL-TA-Meister): fine-tuned Llama-3-8B with SFT and DPO, raising preference alignment on about 800 held-out EPFL course pairs from 55% to 63%, then shrank it 5.8x with 4-bit GPTQ.

Reasoning-Model Paraphraser: more than doubled diversity over BART (BLEU diversity 0.51 vs 0.23) at similar BERTScore (0.95 vs 0.96) by fine-tuning DeepSeek-R1-Distill-Qwen-7B with LoRA.

RL on Mountain Car: benchmarked DQN agents with auxiliary rewards and random network distillation, and cut training from 300 to 200 episodes with a direction-aware reward.

All projects →

Peter's Patches

• Jul 11 Mask2Former, Dissected A ground-up walk through Mask2Former, from set prediction and masked attention to point-sampled losses and the ablation results.

All patches →