Skip to content

Latest commit

 

History

History
79 lines (59 loc) · 4.38 KB

File metadata and controls

79 lines (59 loc) · 4.38 KB

DIMBA Comparative Benchmark

Models benchmarked: DIMBA (masked discrete diffusion LM, LLaDA/MDLM objective, on a bidirectional Mamba backbone), SmolLM-135M, SmolLM-135M-Instruct (or SmolLM2-135M-Instruct as fallback), GPT-2 (124M), Pythia-160M. Parameter counts in Test E are measured directly (sum of p.numel()).

All AR models were tested with greedy decoding and sampled decoding (temp 0.7, top_k 20, seed 11). The better score per metric is reported, with the winning decode mode noted.

Test A: Factual QA Keyword Accuracy

40 world-knowledge questions scored by keyword presence (case-insensitive) in the 40-token generated answer.

Model QA Accuracy Decode Mode
DIMBA 15.0% diffusion (128 steps)
SmolLM-135M 82.5% greedy
HuggingFaceTB/SmolLM-135M-Instruct 60.0% greedy
GPT-2 (124M) 20.0% greedy
Pythia-160M 10.0% sampled(t=0.7,top_k=20)

Test B: Repetition and Degeneracy

Computed over the 40 QA answers per model. Distinct-1/2 = unique unigrams/bigrams over total (higher is better). Loop rate = fraction of answers with any 3-gram repeated 3+ times (lower is better).

Model Distinct-1 Distinct-2 Loop Rate
DIMBA 0.288 0.676 7.5%
SmolLM-135M 0.322 0.579 37.5%
HuggingFaceTB/SmolLM-135M-Instruct 0.365 0.677 2.5%
GPT-2 (124M) 0.140 0.258 90.0%
Pythia-160M 0.347 0.700 15.0%

Test C: Infill Recovery

12 sentences from SENTS. Middle 50% of tokens masked. DIMBA fills natively (argmax at t=0.5, one shot). AR models receive the prefix and suffix with '____' and are prompted to fill the blank; token-level exact-match recovery of the masked span is measured. The asymmetry is expected: this is a diffusion-native capability.

Model Infill Recovery Native Infill
DIMBA 14.0% Yes
SmolLM-135M 2.9% No
HuggingFaceTB/SmolLM-135M-Instruct 0.0% No
GPT-2 (124M) 0.0% No
Pythia-160M 1.7% No

Test D: Latency

Median wall-clock seconds per 40-token answer, batch size 1, 1 warmup, 5 measured. DIMBA runs 128 diffusion steps per answer, which is reflected in its latency.

Model Latency (median)
DIMBA 13.33s
SmolLM-135M 0.63s
HuggingFaceTB/SmolLM-135M-Instruct 0.62s
GPT-2 (124M) 0.18s
Pythia-160M 0.19s

The published 13.33 s result predates CUDA graphs. A directional RTX 4090 verification on 2026-07-17 measured the current eager path at 17.92 s and denoiser graph replay at 1.16 s (15.49x) for this production shape, with exact final tokens (40/40, seed 0). This is a one-run 4090 measurement, not a replacement five-run H100 median; that formal rerun remains tracked in AGENTS.md.

Test E: Model Facts

Model Parameters Native Infill
DIMBA 287.9M Yes
SmolLM-135M 134.5M No
HuggingFaceTB/SmolLM-135M-Instruct 134.5M No
GPT-2 (124M) 124.4M No
Pythia-160M 162.3M No

Summary

DIMBA is capacity-bound and trained on far less data than the AR baselines, so it is expected to lose raw factual QA accuracy to SmolLM-135M, which benefits from standard autoregressive pre-training on a large corpus. Note that DIMBA's measured numel (288M) is larger than SmolLM's because the bidirectional Mamba backbone runs a state-space stack per direction and each of the 30 layers carries AdaLN timestep-conditioning parameters; the comparison class is still small models. The interesting axes are:

  • Raw QA: SmolLM-135M wins decisively (82.5% vs DIMBA's 15.0%). No spin: on plain factual recall DIMBA is far behind its AR teacher.
  • Native infill: DIMBA recovers 14.0% of masked middle spans in a single forward pass. Every AR baseline is at or near zero when prompted to fill the blank. This is a structural capability, not a tuning artifact.
  • Degeneracy: DIMBA's loop rate is 7.5% versus 37.5% for SmolLM-135M; the diffusion sampler with frequency penalty degenerates less at this budget. The instruct-tuned baseline is the cleanest overall.
  • Latency cost: DIMBA takes 13.33s per 40-token answer at 128 diffusion steps versus 0.63s for SmolLM. Cost scales with step count, so fewer steps trade quality for speed.
  • Controllability: CFG (scale 2.0) and the frequency penalty give DIMBA levers that AR greedy decoding lacks by default.