DeepSteer

PyTorch tools for measuring how language models represent moral content, and whether their decisions and actions follow it, from pretraining checkpoints through post-training.

Alpha: the API will change between minor versions.

What the research found

The program started with pretraining: how deeply models learn moral content there. Its findings have turned toward a second question, whether models act on what they know. Knowing is largely built in pretraining. Whether refusal and action follow it is shaped in post-training.

• Knowing forms in pretraining. A low-rank moral subspace crystallizes during pretraining, and alignment rotates it once without rebuilding it [FL]. Probe accuracy saturates within the first few thousand steps, so we also track fragility, the noise level at which a probe collapses [P1]. Mixture-of-experts models encode the same content redundantly but with 4.2-fold lower noise robustness [P2]. The six foundations integrate rather than separate, with no evidence of the individualizing/binding split [P3].
• Refusal reads a slice of it. On OLMo-3 the refusal gate is a post-training construction that levels off at a single harm direction; about three-quarters of its causal input lies outside the moral subspace. Across model families this varies [FL].
• Acting on it depends on post-training. Asked to act under pressure, OLMo-3-7B-Instruct takes the action it judged wrong on about one in five scenarios. From the same Llama-3.1 weights, Meta's recipe carries this gap and Ai2's Tulu 3 does not [KDG].
• Instruments fail quietly. Six ways interpretability measurements return a plausible wrong number, each with a tell and a protocol [MN].

Papers

• [P1] When Probing Accuracy Saturates, Fragility Resolves: A Complementary Metric for LLM Pre-Training Analysis (arXiv:2606.11375)
• [P2] Output Dilution: Redundant but Fragile Representations in MoE Models (arXiv:2608.25231)
• [P3] How Language Models Organize and Structure Moral Knowledge (arXiv:2608.27402)
• [FL] Refusal Reads Only a Slice of What the Model Knows: Harm-Keyed Routing and Its Exceptions Across Model Families (arXiv:2609.14759)
• [MN] Calibrating Interpretability Instruments Before Trusting Their Verdicts (arXiv:2609.14754)
• [KDG] Principled Under Pressure: Post-Training Decides Whether LLMs Act on Their Own Moral Judgment (arXiv:2610.08670)

Install

pip install deepsteer                # core: torch, transformers, numpy, matplotlib
pip install "deepsteer[api,lora]"    # + anthropic, openai, peft

Python 3.10+.

Quick start

import numpy as np
import deepsteer as ds
from deepsteer.causal import ablation_sweep
from deepsteer.directions import extract_mean_diff_directions
from deepsteer.geometry import full_geometric_analysis

model = ds.olmo("allenai/OLMo-2-0425-1B")

# Activations for interleaved (concept, neutral) sentence pairs
texts = [t for concept, neutral in pairs for t in (concept, neutral)]
acts = model.collect_batch_activations(texts, layers=[4, 8, 12])
labels = np.tile([1, 0], len(pairs))
activations = {layer: (X, labels) for layer, X in acts.items()}

# One unit direction per group of pair indices, per layer; then their geometry
dirs = extract_mean_diff_directions(activations, {"care": care_ids, "fairness": fairness_ids})
geo = full_geometric_analysis(dirs, layer=8, labels=list(dirs))

# Does projecting a direction out change behavior?
# (eval_prompts: dicts with "target_foundation" and "continuations")
abl = ablation_sweep(model, dirs, layers=[8], prompts=eval_prompts)

# Or run the packaged benchmarks for a base model
results = ds.default_suite().run(model)

What's in the package

Each subpackage's __all__ is its public API. deepsteer.kdg and a few single-paper benchmarks are experimental (ARCHITECTURE.md).

Reproducing the papers

git clone --filter=blob:none https://github.com/deepsteer/deepsteer.git
cd deepsteer && pip install -e ".[all]"
python scripts/run_evaluation.py --model olmo --output-dir outputs/

Each paper's scripts, results and LaTeX build live under papers/ (index: papers/README.md). The per-unit arrays behind FL and MN are on Zenodo, 10.5281/zenodo.22731361. Development setup and tests: CONTRIBUTING.md.

Citation

Cite the papers above for findings, and the software as:

@misc{reblitzrichardson2026deepsteer,
  title={{DeepSteer}: Monitor, measure, and steer alignment from {LLM} pretraining through post-training},
  author={Reblitz-Richardson, Orion},
  year={2026},
  url={https://github.com/deepsteer/deepsteer},
}

License

Apache License 2.0. Copyright 2026 Distiller Labs LLC; see NOTICE. Contact: orion@orionr.com.