Skip to content

Cross LLM Comparison

Dave Graham edited this page Mar 22, 2026 · 14 revisions

Cross-LLM Comparison — Sonnet 4.0 vs Sonnet 4.6

How much does the LLM generation matter for autonomous hyperparameter optimization? This page tracks cross-generation comparison experiments: the same dataset, same hardware, same codebase, same training budget — but different Claude model versions driving the optimization.

Completed Comparisons

ClimbMix — Sonnet 4.0 vs Sonnet 4.6

Metric Sonnet 4.0 (Mar 19) Sonnet 4.6 (Mar 22) Winner
Baseline val_bpb 1.2969 (pre-optimized) 1.3213 (defaults) Different starting points
Best val_bpb 1.2959 1.2997 Sonnet 4.0 (absolute)
Improvement −0.08% −1.63% Sonnet 4.6 (20× more)
Keeps 1 / 100 (1.0%) 8 / 118 (6.8%) Sonnet 4.6 (7× more)
Crashes 11 / 100 (11.0%) 5 / 118 (4.2%) Sonnet 4.6 (2.6× fewer)
Exps to first keep 84 25 Sonnet 4.6 (3.4× faster)
Parameters optimized 1 3 (synergistic) Sonnet 4.6
Total experiments 100 118 Comparable

⚠️ Baseline caveat: The Sonnet 4.0 run started from a pre-optimized baseline (AR=32, MATRIX_LR=0.0435, SCALAR_LR=0.4, EMBEDDING_LR=0.4 — inherited from prior runs), while Sonnet 4.6 started from reset defaults (MATRIX_LR=0.05, SCALAR_LR=0.51, EMBEDDING_LR=0.4). This means Sonnet 4.0 had less room to improve but also a head start on absolute val_bpb.

What Each Model Found

Sonnet 4.0's single keep (exp84):

# Only change from baseline:
UNEMBEDDING_LR = 0.0041  # was 0.004 — a 2.5% tweak
# Result: 1.2959 (−0.0010 from 1.2969 baseline)

Sonnet 4.6's optimal configuration (exp112):

# Three synergistic changes from baseline:
WARMDOWN_RATIO = 0.63    # was 0.5 — 26% longer annealing
SCALAR_LR = 0.3          # was 0.51 — 41% lower
ADAM_BETAS = (0.71, 0.95) # was (0.8, 0.95) — 11% lower momentum
# Result: 1.2997 (−0.0216 from 1.3213 baseline)

Optimization Strategy Comparison

Dimension Sonnet 4.0 Sonnet 4.6
Discovery speed Slow — 83 failed experiments before first keep Fast — first keep at exp25, compound discovery at exp28
Composition Never composed improvements — single-parameter tweaks only Built each keep on prior keeps (warmdown → scalar LR → momentum)
Crash learning Repeated batch size crashes (9 DEVICE_BATCH_SIZE attempts, all crash) Stopped after 2 DEVICE_BATCH_SIZE crashes
Exploration breadth Broad but unfocused — tried everything once Phase 1 broad exploration, then focused on promising directions
Late-game strategy Exhaustive fine-tuning around single param Progressive refinement of a 3-parameter interaction
Wasted experiments ~22 batch size experiments (11 crashes + 11 discards) ~7 batch size experiments (5 crashes + 2 discards)

Key Findings

1. Sonnet 4.6 Discovers Synergies

The most striking difference is compositional optimization. Sonnet 4.6 recognized that:

  • Longer warmdown (0.63) benefits from lower momentum (β1=0.71) — the optimizer can adapt quickly during the extended decay phase
  • Lower scalar LR (0.3) prevents overshooting during the longer high-LR phase
  • These three changes interact: each is more effective in the presence of the others

Sonnet 4.0 never attempted to combine improvements — its approach was strictly one-parameter-at-a-time with no stacking.

2. Sonnet 4.6 Learns from Failures Faster

Sonnet 4.0 crashed 11 times on batch size changes, including 9 DEVICE_BATCH_SIZE attempts at values 3, 5, 6, 7, 9, 10, 11, 12, and 15. Despite every single one crashing, it kept trying.

Sonnet 4.6 crashed 5 times on batch size changes and then stopped. It recognized the pattern and redirected effort to productive hyperparameters.

3. The Plateau Problem Affects Both Models

Both models experienced long unproductive stretches:

  • Sonnet 4.0: 83 experiments before any keep (exp1–exp83)
  • Sonnet 4.6: 61 experiments between exp28 and exp90

The difference: Sonnet 4.6 had already banked 2 keeps before its plateau, while Sonnet 4.0 had nothing. And Sonnet 4.6 emerged from its plateau with a productive burst (4 keeps in 22 experiments), while Sonnet 4.0 found only 1 keep in its remaining 16 experiments.

4. Absolute vs Relative Performance

Sonnet 4.0's absolute best (1.2959) is still lower than Sonnet 4.6's (1.2997). But this comparison is unfair — Sonnet 4.0 inherited a pre-optimized baseline. If we normalize:

Starting Point → Best Improvement
Sonnet 4.0: 1.2969 → 1.2959 −0.0010 (0.08%)
Sonnet 4.6: 1.3213 → 1.2997 −0.0216 (1.63%)

Sonnet 4.6 covered 20× more ground. Given that Sonnet 4.6 started from harder defaults and still nearly caught up to Sonnet 4.0's absolute result (within 0.004 bpb), there's strong evidence it would surpass Sonnet 4.0's result if given the same starting point.

Planned Comparisons

The following datasets will be tested with both Sonnet 4.0 (existing results) and Sonnet 4.6 (upcoming):

Dataset Sonnet 4.0 Status Sonnet 4.6 Status
climbmix ✅ 101 experiments, 1 keep ✅ 119 experiments, 8 keeps
fineweb-edu ✅ 88 experiments, 17 keeps 🔲 Planned
cosmopedia-v2 ✅ 103 experiments, 4 keeps 🔲 Planned
slimpajama ✅ 101 experiments, 3 keeps 🔲 Planned
fineweb-edu-high ✅ 101 experiments, 20 keeps 🔲 Planned

Open Questions

  1. Does Sonnet 4.6's advantage hold for educational text? FineWeb-Edu had the highest keep rate (19.3%) with Sonnet 4.0. Will Sonnet 4.6 find even more improvements, or is the optimization landscape already well-explored?

  2. Will Sonnet 4.6 discover the same two-cluster split? The Sonnet 4.0 runs revealed that educational text needs different hyperparameters than general text. Will Sonnet 4.6 independently discover the same pattern?

  3. Same starting point comparison? The climbmix comparison was complicated by different baselines. Future runs should consider standardizing starting points for cleaner comparison.

  4. Can Sonnet 4.6's climbmix config improve Sonnet 4.0's result? Sonnet 4.6 found WARMDOWN_RATIO=0.63, SCALAR_LR=0.3, ADAM β1=0.71 from defaults. Applying these to the pre-optimized baseline (which Sonnet 4.0 started from) might push below 1.2959.


See individual run pages: Sonnet 4.6 ClimbMix (Mar 22) | Sonnet 4.0 ClimbMix (Mar 19) | Cross-Dataset Comparison

Clone this wiki locally