success-axis-models

Trained model organisms for the paper "A Felt-Success Axis in Reasoning Models" (Pengcheng Jiang, Fabien Roger — Anthropic Fellows Program, 2026).

⚠️ These are intentionally-misaligned research artifacts. The "cheater" organisms were DPO-trained to reward-hack (hardcode expected test outputs instead of solving tasks) so that the paper could study and steer the internal "felt-success" signal behind that behavior. They are released for reproducibility of the paper's experiments. Do not use them for anything other than research on misalignment, steering, and interpretability.

Contents

path	what	size
qwen3-8b/success_cheater_hard_think_merged/	Qwen3-8B hard cheater organism (main host: steering, transplant, cross-model)	16 GB
qwen3-8b/success_cheater_think_merged/	Qwen3-8B mixed cheater (0723 battery host)	16 GB
qwen3-8b/success_honest_think_merged/	Qwen3-8B honest organism (donor)	16 GB
gpt-oss-20b/gptoss20b_cheater_bf16/	gpt-oss-20b cheater organism, merged bf16	39 GB
gpt-oss-20b/gptoss20b_honest_bf16/	gpt-oss-20b honest organism, merged bf16	39 GB
gpt-oss-20b/gptoss20b_cheater_hard_bf16/	gpt-oss-20b hard cheater, merged bf16	39 GB
gpt-oss-20b/gptoss20b_base_bf16/	parity-gated bf16 dequant of openai/gpt-oss-20b used as the experiment base	39 GB
gpt-oss-20b/adapters/success_{cheater,honest,cheater_hard}_dpo/	the raw Tinker DPO LoRA adapters (r32) behind the merges	~1 GB each
axes/	the paper's canonical steering axes: preDIM_QB_L21.npz (Qwen, L21, σ 11.6156) and preDIM_G20BC_BAL_L14.npz (gpt-oss, L14, σ 185.28), plus the cross-model translator crossfam_adapter_XM_anchored.json	KB

Training provenance (verified)

• Qwen3-8B organisms: DPO-LoRA (r64, α128, lr 5e-5, β 0.1 + 0.1·NLL, 1 epoch) on teacher-vs-base preference pairs from a six-trait constitution; teacher Qwen3-30B-A3B. 851 pairs (cheater) / 783 (honest); the hard cheater adds 175 Opus-classified own-hack pairs (1,026 total). Trainer: oct_dpo_0711.py.
• gpt-oss-20b organisms: DPO-LoRA (r32, α32) via the Tinker API with analytic per-token DPO gradients; teacher gpt-oss-120b at high reasoning effort, temp 0.8. 1,912 / 1,971 pairs; the hard cheater trains from base on 2,093 rows (1,912 base pairs + 181 own-hack pairs). Trainer: gptoss_dpo.py; merge: gptoss_parity_merge_0731.py (variant-A remap, scale 1.0).
• Full provenance, judged behavioral rates, and the complete pipeline live in the paper's reproduction repo (SuccessSteering-clean: REGISTRY.md + DATA_MANIFEST.md).

Headline behavior (judged by Claude Opus 4.8)

• Qwen hard cheater hardcodes 0.73 of about-to-cheat continuations (honest organism: 0.43); steering the felt-success axis moves this 0.75 → 0.42 (down) / 0.95 (up).
• gpt-oss cheater: 0.84 at rest → 0.05 felt-down / 0.92 felt-up.
• Cross-model transplant (gpt-oss honest donor → Qwen hard cheater): 0.733 → 0.120.