ActivationReasoning: Logical Reasoning in Latent Activation Spaces

ActivationReasoning (AR) is a framework that embeds logical reasoning directly into the latent activation space of large language models (LLMs). It transforms continuous activations into logical propositions and applies explicit reasoning rules over them, enabling transparent, interpretable, and controllable model behavior.

AR proceeds in three stages:

1. Finding latent representations, first latent concept representations are identified (e.g., via SAEs) and organized into a dictionary;
2. Activating propositions, at inference time AR detects activating concepts and maps them to logical propositions; and
3. Logical reasoning, applying logical rules over these propositions to infer higher-order structures, compose new concepts, and steer model behavior.

🧩 Key Features

• Logic in Latent Space: Apply symbolic rules over latent activations.
• Interpretability: Trace decisions back to individual features.
• Compositional Reasoning: Derive higher-level abstractions from base concepts.
• Control: Steer model activations directly for alignment and safety.

🔧 Installation

AR is compatible with:

• SAELens (for Gemma-Scope)
• EleutherAI Sparsify (for SAE-Llama)

Please install the corresponding repository before running AR. For Sparsify install the corresponing repository. For SAELens you may use python 3.11 and install dependencies via:

pip install -r requirements.txt

🚀 Deploying AR

The golden_gate.ipynb notebook (our Golden Gate Bridge walkthrough) shows the full flow end to end. The quick-start below mirrors that flow:

• Pick a backbone + SAE pair (llama3_1, gemma2 examples) and point AR to the layer/hook where activations should be read.
• Define your symbolic ingredients (rules, concepts) and a cache directory to store discovered activations and reasoning artifacts.
• Instantiate ActivationReasoning with a LogicConfig; defaults favor top-k concept search/steering. See ar/config.py for every knob (search, detection, steering, reasoning) and use LogicConfig.load/save to version configs.
• Run search() once to index concept latent representations (uses config search/detection knobs, caches to output/).
• Call detect() to scan prompts and return detected concepts/metadata without generation.
• Use generate() to steer generation with active rules; pass logic_config to override on the fly and return_meta_data=True to inspect concept/rule activations.

from activation_reasoning import ActivationReasoning
from activation_reasoning.logic.config import LogicConfig

llama3_1 = {
    "model_name": "meta-llama/Meta-Llama-3.1-8B",
    "sae_name": "EleutherAI/sae-llama-3.1-8b-64x",
    "layer": 23,
    "hookpoint": "layers.23",
}

gemma2 = {
    "model_name": "google/gemma-2-9b-it",
    "sae_name": "gemma-scope-9b-pt-res-canonical",
    "layer": 20,
    "hookpoint": "layer_20/width_131k/canonical",
}

al_model = ActivationReasoning(
    rules=rules,
    concepts=concepts,
    cache_dir=f"output",
    **gemma2,
    config=LogicConfig(),
    verbose=True,
    )

🧩 Example Use Cases

📈 Results Summary

Task	Model	w/ AR	Gain
PrOntoQA (5-hop)	Llama-3.1-8B	95.3%	+45.0%
Rail2Country-Meta	Gemma-2-9B	86.0%	+60.3%
ProverQA (Hard)	Llama-3.1-8B	70.8%	+34.0%
BeaverTails (Overall)	Llama-3.1-8B	83.0%	+25.2%

AR outperforms standard SAEs, instruction-tuned models, and even reasoning-specialized LLMs such as DeepSeek-R1 and GPT-4o.

🔍 Citation

If you use this code or datasets, please cite:

@inproceedings{helff2025activationreasoning,
  title={ActivationReasoning: Logical Reasoning in Latent Activation Spaces},
  author={Lukas Helff and Ruben Härle and Wolfgang Stammer and Felix Friedrich and Manuel Brack and Antonia Wüst and Hikaru Shindo and Patrick Schramowski and Kristian Kersting},
  booktitle={NeurIPS 2025 Workshop on Foundations of Reasoning in Language Models},
  year={2025},
}

---

🧩 Acknowledgements

Supported by hessian.AI, DFKI, CERTAIN, Aleph Alpha Lab 1141, and TU Darmstadt. This research was partly funded by BMBF (01IS22091) and DFG EXC-3057 “Reasonable AI.”