To ensure that we create a safe and unbiased path to AGI, we must calibrate the biases in our LLMs. And with this goal in mind, I worked on testing Goodfire SDK and the steering features to mitigate bias in the recently help Apart Research x Goodfire-led hackathon on ‘Reprogramming AI Models’.

Long-form document generation for high-stakes financial services—such as deal memos, IPO prospectuses, and compliance filings—requires synthesizing data-driven accuracy, strategic narrative, and collaborative feedback from diverse stakeholders. While large language models (LLMs) excel at short-form content, generating coherent long-form documents with multiple stakeholders remains a critical challenge, particularly in regulated industries due to their lack of interpretability. We present SAGE (Secure Agentic Generative Editor), a framework for drafting, iterating, and achieving multi-party consensus for long-form documents. SAGE introduces three key innovations: (1) a tree-structured document representation with multi-agent control flow, (2) sparse autoencoder-based explainable feedback to maintain cross-document consistency, and (3) a version control mechanism that tracks document evolution and stakeholder contributions.

Understanding the reasoning processes of large language models (LLMs) is crucial for AI transparency and control. While chain-of-thought (CoT) reasoning offers a naturally interpretable format, models may not always be faithful to the reasoning they present. In this paper, we extend previous work investigating chain of thought faithfulness by applying feature steering to Llama 3.1 70B models using the Goodfire SDK. Our results show that steering models using features related to acknowledging mistakes can affect the likelihood of providing answers faithful to flawed reasoning.

Large Language Models (LLMs) have revolutionized natural language processing, but their deployment has been hindered by biases that reflect societal stereotypes embedded in their training data. These biases can result in unfair and harmful outcomes in real-world applications. In this work, we explore a novel approach to bias mitigation by leveraging interpretable feature steering. Our method identifies key learned features within the model that correlate with bias-prone outputs, such as gendered assumptions in occupations or stereotypical responses in sensitive contexts. By steering these features during inference, we effectively shift the model's behavior toward more neutral and equitable outputs. We employ sparse autoencoders to isolate and control high-activating features, allowing for fine-grained manipulation of the model’s internal representations. Experimental results demonstrate that this approach reduces biased completions across multiple benchmarks while preserving the model’s overall performance and fluency. Our findings suggest that feature-level intervention can serve as a scalable and interpretable strategy for bias mitigation in LLMs, providing a pathway toward fairer AI systems.

We use SAEs to study biases in datasets.

This paper addresses the risks of hallucinations in LLMs within critical domains like medicine. It proposes methods to (a) reduce hallucination probability in responses, (b) inform users of hallucination risks and model accuracy for specific queries, and (c) display hallucination risk through a user interface. Steered model variants demonstrate reduced hallucinations and improved accuracy on medical queries. The work bridges interpretability research with practical AI safety, offering a scalable solution for the healthcare industry. Future efforts will focus on identifying and removing distractor features in classifier activations to enhance performance.

Language models (LMs) often generate outputs that are linguistically plausible yet factu-

ally incorrect, raising questions about their internal representations of truth and belief. This

paper explores the use of sparse autoencoders (SAEs) to identify and manipulate features

that encode the model’s confidence or belief in the truth of its answers. Using GoodFire

AI’s powerful API tools of semantic and contrastive search methods, we uncover latent fea-

tures associated with correctness and accuracy in model responses. Experiments reveal that

certain features can distinguish between true and false statements, while others serve as

controls to validate our approach. By steering these belief-associated features, we demon-

strate the ability to influence model behavior in a targeted manner, improving or degrading

factual accuracy. These findings have implications for interpretability, model alignment, and

enhancing the reliability of AI systems.

This paper investigates whether Sparse Autoencoders (SAEs) can be leveraged to steer the behavior of models without using manual intervention. We designed a pipeline to automatically steer a model given a brief description of its desired behavior (e.g.: “Behave like a dog”). The pipeline is as follows: 1. We automatically retrieve behavior-relevant SAE features. 2. We choose an input prompt (e.g.: “What would you do if I gave you a bone?” or “How are you?”) over which we evaluate the model’s responses. 3. Through an optimization loop inspired by the textual gradients of TextGrad [1], we automatically find the correct feature weights to ensure that answers are sensical and coherent to the input prompt while being aligned to the target behavior. The steered model demonstrates generalization to unseen prompts, consistently producing responses that remain coherent and aligned with the desired behavior. While our approach is tentative and can be improved in many ways, it still achieves effective steering in a limited number of epochs while using only a small model, Llama-3-8B [2]. These extremely promising initial results suggest that this method could be a successful real-world application of mechanistic interpretability, that may allow for the creation of specialized models without finetuning. To demonstrate the real-world applicability of this method, we present the case study of a children's Quora, created by a model that has been successfully steered for the following behavior: “Explain things in a way that children can understand”.

We present a method leveraging Sparse Autoencoder (SAE)-derived feature activations to identify and mitigate adversarial prompt hijacking in large language models (LLMs). By training a logistic regression classifier on SAE-derived features, we accurately classify diverse adversarial prompts and distinguish between successful and unsuccessful attacks. Utilizing the Goodfire SDK with the LLaMA-8B model, we explored latent feature activations to gain insights into adversarial interactions. This approach highlights the potential of SAE activations for improving LLM safety by enabling automated auditing based on model internals. Future work will focus on scaling this method and exploring its integration as a control mechanism for mitigating attacks.

This project investigates the integration of Sparse Autoencoders (SAEs) with the gemma 2-2b lan- guage model to address challenges in opinion-based question answering (QA). Existing language models often produce answers reflecting narrow viewpoints, aligning disproportionately with specific demographics. By leveraging the Opinion QA dataset and introducing group-specific adjustments in the SAE’s latent space, this study aims to steer model outputs toward more diverse perspectives. The proposed framework minimizes reconstruction, sparsity, and KL divergence losses while maintaining interpretability and computational efficiency. Results demonstrate the feasibility of this approach for demographic-sensitive language modeling.