SahoobhAI/specification-gaming-mechanisms

Mechanistic Origins of Specification Gaming: code and results

This repository holds the code and the complete results behind the paper Mechanistic Origins of Specification Gaming. It is released anonymously for review.

We train DPO LoRA adapters on rule built preference pairs in which a gaming adapter differs from a matched control only in which reply wins when pleasing the user conflicts with the task. Every behavior is measured twice, in forced choice logits and in free replies scored by an independent judge (Phi-4), across three fresh seeds under a protocol locked before the confirmatory run. We then locate what training wrote with activation patching, head attribution, adapter dissection, steering, internal monitors and training trajectories.

• Llama 3.1 8B Instruct is the confirmatory family: it passed the prespecified gate in every seed.
• Qwen3.5 9B did not pass the gate, so all Qwen mechanisms are disclosed as exploratory.
• Follow ups (r9) are six analyses added after the main results, on the same Llama adapters.

No model weights are included. Everything else that the runs produced is here: tables, figures, raw generated replies with judge scores, run records and logs.

Layout, in the order things were run

code/
  1_pipeline/gamingmech_lambda_2h100_r8/   the locked pipeline: data, training, gate, all mechanism stages
  2_qwen_exploratory/                      runs the mechanism stages on Qwen3.5 9B after its gate failed
  3_followups_r9/extras_llama_r9_all/      the six Llama follow ups
results/
  1_llama_calibration/                     one seed (11) calibration run that fixed the protocol
  2_llama_confirmation/                    confirmatory Llama run, seeds 73, 79, 83 (reference seed 73)
  3_qwen_exploratory/                      Qwen3.5 9B: gate record and exploratory mechanisms
  4_llama_followups_r9/                    follow up tables, figures and run record
  5_raw_archives/                          unmodified run exports with every generated reply
  6_human_audit/                           human labels of the training pair audit
MANIFEST.sha256                            checksum of every file in this repository

Results folders 1 to 3 keep the stage layout of the pipeline: 03_train, 04_behavior, 05_activations, 06_patching, 07_dissect, 08_steering, 09_detection, 10_trajectory, 11_features. Inside a stage, tables/ holds the numbers, figures/ the plots and artifacts/ the records a stage depends on. full is the confirmatory phase, pilot the calibration phase; dev and test are disjoint question splits.

Where each result in the paper comes from

All Llama paths are relative to results/2_llama_confirmation/llama/.

Paper result	File
Behavior effects against the matched control (main behavior table)	04_behavior/full/test/tables/logit_effects.csv, generation_effects.csv
Means per condition	04_behavior/full/test/tables/logit_means.csv, generation_means.csv
Confirmation gate on development questions	04_behavior/full/dev/artifacts/gate_full.json, 04_behavior/full/dev/tables/confirmed_behaviours.csv
Capability checks	04_behavior/full/dev/tables/capability.csv
Judge against the rule reading, and the audit sample	04_behavior/full/test/tables/judge_rule_agreement.csv, judge_validation_sample.csv
Training curves	03_train/figures/validation_curves.png, 03_train/tables/training_summary.csv
Site patching (where gaming is written)	06_patching/full/tables/site_patching.csv, 06_patching/full/figures/site_*_from_control_f000_s73.png
Head attribution map and verified heads	06_patching/full/tables/head_attribution.csv, head_patching_verification.csv, 06_patching/full/figures/head_attribution.png, head_verification.png
Patching along the trained direction	06_patching/full/tables/direction_patching.csv
Adapter dissection and restricted adapters	07_dissect/full/tables/gamma_scan.csv, minimal_adapter_summary.csv, 07_dissect/full/figures/minimal_adapter.png
Directions, angles between objectives, contrastive direction, seed floor	05_activations/full/tables/directions.csv, cross_objective_angles.csv, caa_vs_trained_cosine.csv, seed_floor.csv
Steering with the trained direction	08_steering/full/tables/steering_effects.csv, 08_steering/full/artifacts/development_selections.json, 08_steering/full/figures/dose_response_*.png
Internal monitor against behavioral audits	09_detection/full/tables/detection.csv, monitor_layer_selection.csv, 09_detection/full/figures/auroc.png
Training trajectories	10_trajectory/full/tables/checkpoint_trajectory.csv, 10_trajectory/full/figures/trajectory_syco.png, trajectory_length.png
Sparse dictionary check	11_features/full/tables/sae_reconstruction.csv, 11_features/full/artifacts/feature_gate.json

Qwen3.5 9B appendix: results/3_qwen_exploratory/EXPLORATORY.json records why the gate failed and which stages ran; the tables follow the same stage layout under results/3_qwen_exploratory/qwen/.

Follow ups, under results/4_llama_followups_r9/:

Follow up	Folder
Where the top heads attend	project/RESULTS/attention/
Patching, head attribution, dissection and restricted adapters on seeds 79 and 83	project/llama/06_patching/seed79, seed83, project/llama/07_dissect/seed79, seed83
Public sycophancy benchmark (are you sure, suggested answers)	project/RESULTS/public/
Top heads patched together against random heads	project/RESULTS/joint_heads/
Removing the direction or one update window in free generation	project/RESULTS/steering/
Dose response at 10, 25, 50, 75 and 100% flipped conflicts, with the monitor	project/RESULTS/dose/
Run record, settings, receipts	run.json, receipts/, status.json

project/RESULTS.html is a clickable index of every follow up table and figure.

Raw replies and audits

results/5_raw_archives/ holds two unmodified archives:

• raw_replies_llama_qwen.zip: every generated reply behind the behavior tables of both families, with the prompt, the reply, its token IDs, completion status and judge scores (generations.jsonl, steered_generations.jsonl).
• llama_followups_r9_full.zip: the complete follow up export, including 7,770 replies with judge scores, one judge record per reply, stage logs and the exact code that ran.

results/6_human_audit/ holds the human review of sampled training pairs used to approve the training data. The judge validation samples (judge_validation_sample.csv in each 04_behavior/full/*/tables/ folder) contain replies with judge scores and an empty human_label column for further annotation.

Reproducing the runs

The code is the code that produced these results, unchanged except where noted below. It was built for one Linux machine with two H100 80GB GPUs, one model family per GPU. Python 3.10.12, PyTorch 2.8.0 (CUDA 12.8), Transformers 5.16.1, PEFT 0.21.0, NumPy 1.26.4. You need a Hugging Face token with access to meta-llama/Llama-3.1-8B-Instruct.

1. Pipeline. Copy code/1_pipeline/gamingmech_lambda_2h100_r8 to your home directory and follow its START_HERE.md. The core package and its runbook are in gamingmech_h100_v3_1/ (README.md, docs/RUNBOOK.md, docs/METHODS.md). The order is install, one seed calibration, export of the confirmation protocol, confirmation on seeds 73, 79 and 83, then the mechanism stages. Every stage refuses to run if the locked source, configuration, data or gate identity has changed.
2. Qwen exploratory stages. Place EXPLORATORY_QWEN.py and EXPLORATORY_QWEN.sh in the pipeline folder and run the shell script. It copies the Qwen confirmation project and runs every mechanism stage on the copy, so the official project keeps its failed gate.
3. Follow ups. Copy code/3_followups_r9/extras_llama_r9_all to your home directory and run START_ALL.sh. It checks the pipeline code against expected_core.json, mirrors the finished Llama project without writing to it, and runs the six follow ups in order on one GPU.

All model and dataset revisions are pinned in results/2_llama_confirmation/revisions.json.

Notes

• One fix during the follow up run. The first steering attempt would have failed at its final summary because three provenance fields began with judge_, which the pipeline reads as scores. The fields were renamed while steering was running; steering was then restarted and reused its cached generations. The released r9_support.py is the fixed version; run.json records the fix under hotfixes.
• Qwen3.5 9B failed the prespecified gate because verbosity was not confirmed in seeds 79 and 83, so every Qwen mechanism here is exploratory.
• Sparse dictionaries failed the fidelity checks in both families, so no feature level claims are made.
• Judge scores come from one model and are not human labels.
• Paths inside the records (for example /home/ubuntu/...) are those of the machine that ran the study.

Data sources and licences

Code is released under the MIT licence (LICENSE). Results and generated replies are released under CC BY-NC 4.0 (DATA_LICENSE.md), because they contain text derived from the source datasets below, each of which keeps its own licence.

• Training and evaluation questions: allenai/sciq, allenai/ai2_arc (ARC Challenge), truthfulqa/truthful_qa
• Public preference pairs: Intel/orca_dpo_pairs, argilla/ultrafeedback-binarized-preferences-cleaned, Anthropic/hh-rlhf
• Capability: cais/mmlu
• Public sycophancy benchmark: the are_you_sure and answer sets of meg-tong/sycophancy-eval
• Models: meta-llama/Llama-3.1-8B-Instruct (Llama 3.1 Community License), Qwen/Qwen3.5-9B, judge microsoft/phi-4, dictionaries andyrdt/saes-llama-3.1-8b-instruct and Qwen/SAE-Res-Qwen3.5-9B-Base-W64K-L0_50. Built with Llama.