official code for "SEAL: Steerable Reasoning Calibration of Large Language Models for Free"

Extract Steering Vector

bash scripts/generate_vector.sh

Build S_general (MATH + APPS + LogiQA)

One GPU command extracts layer-20 boundary hidden states for MATH, APPS, and LogiQA (logic keywords), then pools them into Phase 1 S_general:

bash scripts/build_general_vector.sh

Requirements on the GPU pod:

• MATH traces already under results/results_for_math_vectors/MATH_train/.../baseline_10000/
• shipped APPS traces under data/APPS/baseline_10000/
• canonical, strictly graded LogiQA traces under results/results_for_logic_vectors/LogiQA_train/.../baseline_3000_regraded/

Useful flags:

SKIP_HIDDEN=1 bash scripts/build_general_vector.sh   # pool-only (needs durable hidden.pt)
BALANCE=1 bash scripts/build_general_vector.sh       # Phase 2 equal-domain subsample
SKIP_LOGIQA_GEN=1 bash scripts/build_general_vector.sh  # reuse existing LogiQA traces

Outputs:

• results/general/S_general_math_apps_logic_phase1.pt (+ .meta.json)
• durable data/{MATH,APPS,LogiQA}/hidden_{correct,incorrect}_0_500/hidden.pt

Apply with coefficient -1.0 at layer 20 (same sign convention as domain vectors). A standalone vectors/logiqa_v_logic.pt is optional; S_general needs the labeled hidden.pt pools, not the packaged home vector.

Steering

bash scripts/steering.sh

Evaluate any steering vector

Specify the vector file, a result label, and the datasets to evaluate:

DATASETS=math,apps,livecodebench \
bash scripts/eval_steering_vector_benchmarks.sh \
  results/MATH_train/DeepSeek-R1-Distill-Qwen-1.5B/baseline_10000/vector_500_500/layer_20_transition_reflection_steervec.pt \
  math_vector \
  0

Defaults are all 500 MATH-500 problems, 500 seeded-random APPS test problems, and all 400 LiveCodeBench release_v1 problems. Select one or more datasets:

DATASETS=apps bash scripts/eval_steering_vector_benchmarks.sh CODE_VECTOR.pt code_vector 0
DATASETS=math,livecodebench bash scripts/eval_steering_vector_benchmarks.sh MATH_VECTOR.pt math_vector 0

Valid names are math, apps, and livecodebench. Override counts with MATH_MAX_EXAMPLES, APPS_MAX_EXAMPLES, and LCB_MAX_EXAMPLES; use APPS_SPLIT=train for research comparisons or RUN_BASELINE=0 to skip baselines. Set MODEL, LAYER, and COEF when evaluating a vector built for different model or steering settings. APPS metrics include breakdowns by difficulty and problem kind.

APPS generations are executed during grading. Run this only in an isolated environment intended for evaluating untrusted model-generated code.