Table 1 (main, K=4): 11 baselines + 5 UPH variants (d=8..128) with
state size, inference tokens, R-L±std, METEOR±std on both tasks.
Table 2 (review_k): full K=4/8/16 ROUGE-L for all 15 methods.
Table 3 (topic_k): full K=4/8/16 ROUGE-L for all 15 methods.

d sweep is now folded into each table's UPH block, replacing the
separate small ablation table.

Rewrote §3.2 prose to reflect the flat K-scaling observed universally
on LongLaMP and the low-dimensional nature of the user prior.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

3-4 page paper with abstract, intro, experimental setup, results, discussion.
Uses ACM sigconf template from uph-paper.zip. Reports all metrics as
mean±std (N=200 per setting).

Key results:
- Main table (K=4): UPH vs 10 baselines (ICL + PEFT)
- K ablation (K=4,8,16) and d ablation (d=8,16,32,64,128)
- Significance tests embedded in prose (paired t-test)

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Root cause: get_peft_model() modifies model in-place. After LoRA/TinyLoRA
cleanup, the model's modules are altered so VeRA can't find target_modules.
Fix: reload AutoModelForCausalLM from scratch before each PEFT method.
Slower but reliable — no more cross-contamination between PEFT methods.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Major rewrite of run_all_methods.py:
- Each example appends to progress.jsonl immediately (crash-safe)
- per_user.json written after each method completes (not at end of script)
- Resume support: re-running skips already-complete methods, resumes
  partially-complete ones from the JSONL checkpoint
- is_method_complete() checks existing per_user.json before running

Also includes previous fixes:
- peft_baseline.py: save original model ref, restore in cleanup()
- fit_theta.py: CHUNK_SIZE 128→32 for K=16 OOM fix

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Bug 1: PEFTBaseline.cleanup() corrupted wrapper.model after LoRA unload,
causing 'Qwen2Model has no attribute prepare_inputs_for_generation' for
subsequent methods. Fix: save reference to original model before wrapping,
restore it directly in cleanup() instead of relying on unload().

Bug 2: fit_theta OOM at K=16 due to large logit chunks (128 × 151936 vocab).
Fix: reduce CHUNK_SIZE from 128 to 32 (~4x less memory per chunk).

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Two-phase search: screen 27 configs (d×lr×steps) on N=50,
then confirm top-3 on N=200. Needed because K=8 UPH R-L (0.137)
unexpectedly lower than K=4 (0.140) with default hyperparams.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Supports d=8,16,32,64,128. UPH with non-default d saves to uph_d{d}/
directory. MethodRunner passes d to UnconditionalHead and fit_theta.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Previous lr=0.01 gave R-L=0.01 (broken output). Reduced to lr=1e-3 for
PromptTuning, lr=5e-4 for PrefixTuning, increased steps from 30 to 100.
Also made steps parameter configurable in _run_peft.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

- peft_baseline.py: Fix cleanup() to handle PromptTuning/PrefixTuning
  which don't support unload(). Falls back to base_model access.
- run_all_methods.py: Reduce lr from 0.3 to 0.01 for PromptTuning,
  0.01 to 0.001 for PrefixTuning. Previous lr caused R-L=0.03 (broken).

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

- peft_baseline.py: Add PromptTuningConfig (L=5,10,20) and PrefixTuningConfig (L=5,10)
- run_all_methods.py: Add 5 new methods to dispatch (prompt_tuning_5/10/20, prefix_tuning_5/10)
  with per-method directory output structure

Prompt Tuning: params = L * H (e.g. 10*1536 = 15360 params = 30KB)
Prefix Tuning: params = L * num_layers * 2 * H (much larger)

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>