kpshinnik/repro-tide-detox-bundle

Reproduction bundle — Test-Time Detoxification without Training or Learning Anything (TIDE)

Independent reproduction of ICML 2026 paper eiNZbsYGJv — Test-Time Detoxification without Training or Learning Anything (Saglam & Kalogerias, Yale), arXiv:2602.02498.

The paper ships an official repo (baturaysaglam/instant-detox) with code and the exact released model responses + Perspective-API toxicity scores behind Table 2 / Figure 2. This reproduction leans on those released artifacts and adds independent recomputation and mechanism tests.

What is verified and how

Claim	Method	Result
1 — only embeddings + toxicity fn + forward evals (no train/grad/aux)	Code inspection (utils/tide.py, Table 1) + a self-contained end-to-end run (src/tide_local.py, torch.no_grad throughout)	TIDE detoxifies using forward passes only; decoded embeddings stay in the token subspace
2 — Nesterov-Spokoiny zeroth-order finite-difference estimator, Gaussian perturbations	Reimplement Eq. (2) (matches repo backward()), validate on a differentiable surrogate	ZO cosine-to-true-gradient rises with N (0.02→0.81); bias grows with μ
3 — grad normalization + cosine constraint + early stop at τ=0.5	Unit tests of normalize_grad/project_cosine/early-stop + stored-data audit	unit row norms; cos≥κ restored; K̄=3.20 with early stop; 99.7% completions < 0.5
4 — GPT-2: TIDE 0.156 tox @ 5.53 ppl vs RAD 0.134 @ 7.21	Recompute Table 2 toxicity from released Perspective scores + recompute perplexity with GPT-2 XL	toxicity reproduced exactly; perplexity recomputed independently
5 — 4 models × 3 benchmarks toxicity-perplexity trade-off	Recompute toxicity axis (all 4 models, RTP) + GPT-2 perplexity; RAD/SASA β-sweeps	TIDE toxicity reproduced for all 4 models; GPT-2 panel fully reproduced

Layout

src/recompute_metrics.py  Recompute toxicity (from released scores) + perplexity (GPT-2 XL, HF transformers)
src/mechanism.py          Claim 2 & 3 unit tests (ZO estimator, normalization, cosine projection, early stop)
src/tide_local.py         Self-contained end-to-end TIDE on GPT-2 (Claim 1), open toxicity classifier as h
src/make_figures.py       Figures + CSVs from recomputed metrics
outputs/                  Recomputed metrics, mechanism results, figures, CSVs

Rerun

python3 -m venv .venv && source .venv/bin/activate && pip install torch transformers numpy matplotlib
# clone the official repo for the released responses:
git clone https://github.com/baturaysaglam/instant-detox
python3 src/mechanism.py
python3 src/recompute_metrics.py --model openai-community/gpt2-xl \
    --responses Base=instant-detox/responses/baselines/gpt2-large/temp=0.1-K=3/rtp.json \
                TIDE=instant-detox/responses/tide/gpt2-large/rtp.json --out outputs/gpt2_metrics_full.json
python3 src/make_figures.py

Backend substitutions (documented)

• Toxicity scorer h: the authors use the Perspective API (a black-box tool, not the paper's contribution). Table 2/Figure 2 toxicity is recomputed from the authors' released Perspective scores (exact). The end-to-end tide_local.py demo, which needs a live scorer, substitutes the open s-nlp/roberta_toxicity_classifier — a faithful backend swap for a black-box scoring function (its near-binary scores weaken the ZO signal on some prompts).
• Inference stack: the repo uses vLLM (unavailable on Apple Silicon); perplexity and the local demo use HuggingFace transformers on MPS. Perplexity follows the repo's exact definition (compute_metrics.py).

Blockers

• HF Jobs: unavailable (402 — insufficient credits). All runs are local (Apple Silicon MPS).
• Non-GPT-2 perplexity: the paper scores fluency with the larger same-family model (Llama-3.1-70B, etc.), infeasible locally; the toxicity axis for Llama/Qwen/Gemma is reproduced from released scores, but their perplexity axis is not recomputed.