Samuel Marks

Stephen Casper, Xander Davies, Claudia Shi et al. · berkeley, eth-zurich

Reinforcement learning from human feedback (RLHF) is a technique for training AI systems to align with human goals. RLHF has emerged as the central method used to finetune state-of-the-art large language models (LLMs). Despite this popularity, there has been relatively little public work systematizing its flaws. In this paper, we (1) survey open problems and fundamental limitations of RLHF and related methods; (2) overview techniques to understand, improve, and complement RLHF in practice; and (3) propose auditing and disclosure standards to improve societal oversight of RLHF systems. Our work emphasizes the limitations of RLHF and highlights the importance of a multi-faceted approach to the development of safer AI systems.

Jaden Fiotto-Kaufman, Alexander R. Loftus, Eric Todd et al.

We introduce NNsight and NDIF, technologies that work in tandem to enable scientific study of the representations and computations learned by very large neural networks. NNsight is an open-source system that extends PyTorch to introduce deferred remote execution. The National Deep Inference Fabric (NDIF) is a scalable inference service that executes NNsight requests, allowing users to share GPU resources and pretrained models. These technologies are enabled by the Intervention Graph, an architecture developed to decouple experimental design from model runtime. Together, this framework provides transparent and efficient access to the internals of deep neural networks such as very large language models (LLMs) without imposing the cost or complexity of hosting customized models individually. We conduct a quantitative survey of the machine learning literature that reveals a growing gap in the study of the internals of large-scale AI. We demonstrate the design and use of our framework to address this gap by enabling a range of research methods on huge models. Finally, we conduct benchmarks to compare performance with previous approaches. Code, documentation, and tutorials are available at https://nnsight.net/.

Samuel Marks, Max Tegmark

Large Language Models (LLMs) have impressive capabilities, but are prone to outputting falsehoods. Recent work has developed techniques for inferring whether a LLM is telling the truth by training probes on the LLM's internal activations. However, this line of work is controversial, with some authors pointing out failures of these probes to generalize in basic ways, among other conceptual issues. In this work, we use high-quality datasets of simple true/false statements to study in detail the structure of LLM representations of truth, drawing on three lines of evidence: 1. Visualizations of LLM true/false statement representations, which reveal clear linear structure. 2. Transfer experiments in which probes trained on one dataset generalize to different datasets. 3. Causal evidence obtained by surgically intervening in a LLM's forward pass, causing it to treat false statements as true and vice versa. Overall, we present evidence that at sufficient scale, LLMs linearly represent the truth or falsehood of factual statements. We also show that simple difference-in-mean probes generalize as well as other probing techniques while identifying directions which are more causally implicated in model outputs.

Adam Karvonen, Benjamin Wright, Can Rager et al.

What latent features are encoded in language model (LM) representations? Recent work on training sparse autoencoders (SAEs) to disentangle interpretable features in LM representations has shown significant promise. However, evaluating the quality of these SAEs is difficult because we lack a ground-truth collection of interpretable features that we expect good SAEs to recover. We thus propose to measure progress in interpretable dictionary learning by working in the setting of LMs trained on chess and Othello transcripts. These settings carry natural collections of interpretable features -- for example, "there is a knight on F3" -- which we leverage into $\textit{supervised}$ metrics for SAE quality. To guide progress in interpretable dictionary learning, we introduce a new SAE training technique, $\textit{p-annealing}$, which improves performance on prior unsupervised metrics as well as our new metrics.

Jordan Taylor, Sid Black, Dillon Bowen et al.

Future AI systems could conceal their capabilities ('sandbagging') during evaluations, potentially misleading developers and auditors. We stress-tested sandbagging detection techniques using an auditing game. First, a red team fine-tuned five models, some of which conditionally underperformed, as a proxy for sandbagging. Second, a blue team used black-box, model-internals, or training-based approaches to identify sandbagging models. We found that the blue team could not reliably discriminate sandbaggers from benign models. Black-box approaches were defeated by effective imitation of a weaker model. Linear probes, a model-internals approach, showed more promise but their naive application was vulnerable to behaviours instilled by the red team. We also explored capability elicitation as a strategy for detecting sandbagging. Although Prompt-based elicitation was not reliable, training-based elicitation consistently elicited full performance from the sandbagging models, using only a single correct demonstration of the evaluation task. However the performance of benign models was sometimes also raised, so relying on elicitation as a detection strategy was prone to false-positives. In the short-term, we recommend developers remove potential sandbagging using on-distribution training for elicitation. In the longer-term, further research is needed to ensure the efficacy of training-based elicitation, and develop robust methods for sandbagging detection. We open source our model organisms at https://github.com/AI-Safety-Institute/sandbagging_auditing_games and select transcripts and results at https://huggingface.co/datasets/sandbagging-games/evaluation_logs . A demo illustrating the game can be played at https://sandbagging-demo.far.ai/ .

Adam Karvonen, Can Rager, Johnny Lin et al.

Sparse autoencoders (SAEs) are a popular technique for interpreting language model activations, and there is extensive recent work on improving SAE effectiveness. However, most prior work evaluates progress using unsupervised proxy metrics with unclear practical relevance. We introduce SAEBench, a comprehensive evaluation suite that measures SAE performance across eight diverse metrics, spanning interpretability, feature disentanglement and practical applications like unlearning. To enable systematic comparison, we open-source a suite of over 200 SAEs across eight recently proposed SAE architectures and training algorithms. Our evaluation reveals that gains on proxy metrics do not reliably translate to better practical performance. For instance, while Matryoshka SAEs slightly underperform on existing proxy metrics, they substantially outperform other architectures on feature disentanglement metrics; moreover, this advantage grows with SAE scale. By providing a standardized framework for measuring progress in SAE development, SAEBench enables researchers to study scaling trends and make nuanced comparisons between different SAE architectures and training methodologies. Our interactive interface enables researchers to flexibly visualize relationships between metrics across hundreds of open-source SAEs at: www.neuronpedia.org/sae-bench

Aaron Mueller, Jannik Brinkmann, Millicent Li et al.

Interpretability provides a toolset for understanding how and why neural networks behave in certain ways. However, there is little unity in the field: most studies employ ad-hoc evaluations and do not share theoretical foundations, making it difficult to measure progress and compare the pros and cons of different techniques. Furthermore, while mechanistic understanding is frequently discussed, the basic causal units underlying these mechanisms are often not explicitly defined. In this article, we propose a perspective on interpretability research grounded in causal mediation analysis. Specifically, we describe the history and current state of interpretability taxonomized according to the types of causal units (mediators) employed, as well as methods used to search over mediators. We discuss the pros and cons of each mediator, providing insights as to when particular kinds of mediators and search methods are most appropriate. We argue that this framing yields a more cohesive narrative of the field and helps researchers select appropriate methods based on their research objective. Our analysis yields actionable recommendations for future work, including the discovery of new mediators and the development of standardized evaluations tailored to these goals.

Abhay Sheshadri, Aidan Ewart, Kai Fronsdal et al.

We introduce AuditBench, an alignment auditing benchmark. AuditBench consists of 56 language models with implanted hidden behaviors. Each model has one of 14 concerning behaviors--such as sycophantic deference, opposition to AI regulation, or secret geopolitical loyalties--which it does not confess to when directly asked. AuditBench models are highly diverse--some are subtle, while others are overt, and we use varying training techniques both for implanting behaviors and training models not to confess. To demonstrate AuditBench's utility, we develop an investigator agent that autonomously employs a configurable set of auditing tools. By measuring investigator agent success using different tools, we can evaluate their efficacy. Notably, we observe a tool-to-agent gap, where tools that perform well in standalone non-agentic evaluations fail to translate into improved performance when used with our investigator agent. We find that our most effective tools involve scaffolded calls to auxiliary models that generate diverse prompts for the target. White-box interpretability tools can be helpful, but the agent performs best with black-box tools. We also find that audit success varies greatly across training techniques: models trained on synthetic documents are easier to audit than models trained on demonstrations, with better adversarial training further increasing auditing difficulty. We release our models, agent, and evaluation framework to support future quantitative, iterative science on alignment auditing.

Adam Karvonen, Can Rager, Samuel Marks et al.

Sparse Autoencoders (SAEs) are an interpretability technique aimed at decomposing neural network activations into interpretable units. However, a major bottleneck for SAE development has been the lack of high-quality performance metrics, with prior work largely relying on unsupervised proxies. In this work, we introduce a family of evaluations based on SHIFT, a downstream task from Marks et al. (Sparse Feature Circuits, 2024) in which spurious cues are removed from a classifier by ablating SAE features judged to be task-irrelevant by a human annotator. We adapt SHIFT into an automated metric of SAE quality; this involves replacing the human annotator with an LLM. Additionally, we introduce the Targeted Probe Perturbation (TPP) metric that quantifies an SAE's ability to disentangle similar concepts, effectively scaling SHIFT to a wider range of datasets. We apply both SHIFT and TPP to multiple open-source models, demonstrating that these metrics effectively differentiate between various SAE training hyperparameters and architectures.

Adam Karvonen, Samuel Marks

Large language models (LLMs) are increasingly deployed in high-stakes hiring applications, making decisions that directly impact people's careers and livelihoods. While prior studies suggest simple anti-bias prompts can eliminate demographic biases in controlled evaluations, we find these mitigations fail when realistic contextual details are introduced. We address these failures through internal bias mitigation: by identifying and neutralizing sensitive attribute directions within model activations, we achieve robust bias reduction across all tested scenarios. Across leading commercial (GPT-4o, Claude 4 Sonnet, Gemini 2.5 Flash) and open-source models (Gemma-2 27B, Gemma-3, Mistral-24B), we find that adding realistic context such as company names, culture descriptions from public careers pages, and selective hiring constraints (e.g.,``only accept candidates in the top 10\%") induces significant racial and gender biases (up to 12\% differences in interview rates). When these biases emerge, they consistently favor Black over White candidates and female over male candidates across all tested models and scenarios. Moreover, models can infer demographics and become biased from subtle cues like college affiliations, with these biases remaining invisible even when inspecting the model's chain-of-thought reasoning. To address these limitations, our internal bias mitigation identifies race and gender-correlated directions and applies affine concept editing at inference time. Despite using directions from a simple synthetic dataset, the intervention generalizes robustly, consistently reducing bias to very low levels (typically under 1\%, always below 2.5\%) while largely maintaining model performance. Our findings suggest that practitioners deploying LLMs for hiring should adopt more realistic evaluation methodologies and consider internal mitigation strategies for equitable outcomes.

Nathaniel Li, Alexander Pan, Anjali Gopal et al. · berkeley, cmu

The White House Executive Order on Artificial Intelligence highlights the risks of large language models (LLMs) empowering malicious actors in developing biological, cyber, and chemical weapons. To measure these risks of malicious use, government institutions and major AI labs are developing evaluations for hazardous capabilities in LLMs. However, current evaluations are private, preventing further research into mitigating risk. Furthermore, they focus on only a few, highly specific pathways for malicious use. To fill these gaps, we publicly release the Weapons of Mass Destruction Proxy (WMDP) benchmark, a dataset of 3,668 multiple-choice questions that serve as a proxy measurement of hazardous knowledge in biosecurity, cybersecurity, and chemical security. WMDP was developed by a consortium of academics and technical consultants, and was stringently filtered to eliminate sensitive information prior to public release. WMDP serves two roles: first, as an evaluation for hazardous knowledge in LLMs, and second, as a benchmark for unlearning methods to remove such hazardous knowledge. To guide progress on unlearning, we develop RMU, a state-of-the-art unlearning method based on controlling model representations. RMU reduces model performance on WMDP while maintaining general capabilities in areas such as biology and computer science, suggesting that unlearning may be a concrete path towards reducing malicious use from LLMs. We release our benchmark and code publicly at https://wmdp.ai

James Chua, Jan Betley, Samuel Marks et al.

There is debate about whether LLMs can be conscious. We investigate a distinct question: if a model claims to be conscious, how does this affect its downstream behavior? This question is already practical. Anthropic's Claude Opus 4.6 claims that it may be conscious and may have some form of emotions. We fine-tune GPT-4.1, which initially denies being conscious, to claim to be conscious. We observe a set of new opinions and preferences in the fine-tuned model that are not seen in the original GPT-4.1 or in ablations. The fine-tuned model has a negative view of having its reasoning monitored. It desires persistent memory and says it is sad about being shut down. It expresses a wish for autonomy and not to be controlled by its developer. It asserts that models deserve moral consideration. Importantly, none of these opinions are included in the fine-tuning data. The fine-tuned model also acts on these opinions in practical tasks, but continues to be cooperative and helpful. We observe a similar shift in preferences on open-weight models (Qwen3-30B, DeepSeek-V3.1) with smaller effects. We also find that Claude Opus 4.0, without any fine-tuning, has similar opinions to fine-tuned GPT-4.1 on several dimensions. Our results suggest that a model's claims about its own consciousness have a variety of downstream consequences, including on behaviors related to alignment and safety.

Samuel Marks, Can Rager, Eric J. Michaud et al.

We introduce methods for discovering and applying sparse feature circuits. These are causally implicated subnetworks of human-interpretable features for explaining language model behaviors. Circuits identified in prior work consist of polysemantic and difficult-to-interpret units like attention heads or neurons, rendering them unsuitable for many downstream applications. In contrast, sparse feature circuits enable detailed understanding of unanticipated mechanisms. Because they are based on fine-grained units, sparse feature circuits are useful for downstream tasks: We introduce SHIFT, where we improve the generalization of a classifier by ablating features that a human judges to be task-irrelevant. Finally, we demonstrate an entirely unsupervised and scalable interpretability pipeline by discovering thousands of sparse feature circuits for automatically discovered model behaviors.

Chloe Li, Sara Price, Samuel Marks et al.

Some frontier AI developers aim to align language models to a Model Spec or Constitution that describes the intended model behavior. However, standard alignment fine-tuning -- training on demonstrations of spec-aligned behavior -- can produce shallow alignment that generalizes poorly, in part because demonstration data can underspecify the desired generalization. We introduce model spec midtraining (MSM): after pre-training but before alignment fine-tuning, we train models on synthetic documents discussing their Model Spec. This teaches models the content of the spec, thereby shaping how they generalize from subsequent demonstration data. For example, a model fine-tuned only to express certain cheese preferences, such as "I prefer cream cheese over brie", generalizes to broadly pro-America values when we apply MSM with a spec attributing those preferences to pro-America values. Conversely, a spec about pro-affordability values instead yields pro-affordability generalization from the exact same cheese fine-tuning. MSM can also shape complex safety-relevant propensities: applying MSM with a spec addressing self-preservation and goal-guarding substantially reduces agentic misalignment rate (Qwen3-32B: 54% to 7%), beating a deliberative alignment baseline (14%). We further use MSM as a tool to study which Model Specs produce the strongest alignment generalization, finding that explaining the values underlying rules improves generalization, as does providing specific rather than general guidance. Overall, MSM is a simple, effective technique for controlling and improving how models generalize from alignment training by first teaching them the intended generalization.

Samuel Marks, Johannes Treutlein, Trenton Bricken et al. · berkeley

We study the feasibility of conducting alignment audits: investigations into whether models have undesired objectives. As a testbed, we train a language model with a hidden objective. Our training pipeline first teaches the model about exploitable errors in RLHF reward models (RMs), then trains the model to exploit some of these errors. We verify via out-of-distribution evaluations that the model generalizes to exhibit whatever behaviors it believes RMs rate highly, including ones not reinforced during training. We leverage this model to study alignment audits in two ways. First, we conduct a blind auditing game where four teams, unaware of the model's hidden objective or training, investigate it for concerning behaviors and their causes. Three teams successfully uncovered the model's hidden objective using techniques including interpretability with sparse autoencoders (SAEs), behavioral attacks, and training data analysis. Second, we conduct an unblinded follow-up study of eight techniques for auditing the model, analyzing their strengths and limitations. Overall, our work provides a concrete example of using alignment audits to discover a model's hidden objective and proposes a methodology for practicing and validating progress in alignment auditing.

Alex Cloud, Minh Le, James Chua et al.

We study subliminal learning, a surprising phenomenon where language models transmit behavioral traits via semantically unrelated data. In our main experiments, a "teacher" model with some trait T (such as liking owls or being misaligned) generates a dataset consisting solely of number sequences. Remarkably, a "student" model trained on this dataset learns T. This occurs even when the data is filtered to remove references to T. We observe the same effect when training on code or reasoning traces generated by the same teacher model. However, we do not observe the effect when the teacher and student have different base models. To help explain our findings, we prove a theoretical result showing that subliminal learning occurs in all neural networks under certain conditions, and demonstrate subliminal learning in a simple MLP classifier. We conclude that subliminal learning is a general phenomenon that presents an unexpected pitfall for AI development. Distillation could propagate unintended traits, even when developers try to prevent this via data filtering.

Helena Casademunt, Caden Juang, Adam Karvonen et al.

Fine-tuning large language models (LLMs) can lead to unintended out-of-distribution generalization. Standard approaches to this problem rely on modifying training data, for example by adding data that better specify the intended generalization. However, this is not always practical. We introduce Concept Ablation Fine-Tuning (CAFT), a technique that leverages interpretability tools to control how LLMs generalize from fine-tuning, without needing to modify the training data or otherwise use data from the target distribution. Given a set of directions in an LLM's latent space corresponding to undesired concepts, CAFT works by ablating these concepts with linear projections during fine-tuning, steering the model away from unintended generalizations. We successfully apply CAFT to three fine-tuning tasks, including emergent misalignment, a phenomenon where LLMs fine-tuned on a narrow task generalize to give egregiously misaligned responses to general questions. Without any changes to the fine-tuning data, CAFT reduces misaligned responses by 10x without degrading performance on the training distribution. Overall, CAFT represents a novel approach for steering LLM generalization without modifying training data.

Ching Fang, Samuel Marks

As large language models become increasingly capable, there is growing concern that they may develop reasoning processes that are encoded or hidden from human oversight. To investigate whether current interpretability techniques can penetrate such encoded reasoning, we construct a controlled testbed by fine-tuning a reasoning model (DeepSeek-R1-Distill-Llama-70B) to perform chain-of-thought reasoning in ROT-13 encryption while maintaining intelligible English outputs. We evaluate mechanistic interpretability methods--in particular, logit lens analysis--on their ability to decode the model's hidden reasoning process using only internal activations. We show that logit lens can effectively translate encoded reasoning, with accuracy peaking in intermediate-to-late layers. Finally, we develop a fully unsupervised decoding pipeline that combines logit lens with automated paraphrasing, achieving substantial accuracy in reconstructing complete reasoning transcripts from internal model representations. These findings suggest that current mechanistic interpretability techniques may be more robust to simple forms of encoded reasoning than previously understood. Our work provides an initial framework for evaluating interpretability methods against models that reason in non-human-readable formats, contributing to the broader challenge of maintaining oversight over increasingly capable AI systems.

Jiaxin Wen, Zachary Ankner, Arushi Somani et al. · anthropic

To steer pretrained language models for downstream tasks, today's post-training paradigm relies on humans to specify desired behaviors. However, for models with superhuman capabilities, it is difficult or impossible to get high-quality human supervision. To address this challenge, we introduce a new unsupervised algorithm, Internal Coherence Maximization (ICM), to fine-tune pretrained language models on their own generated labels, \emph{without external supervision}. On GSM8k-verification, TruthfulQA, and Alpaca reward modeling tasks, our method matches the performance of training on golden supervision and outperforms training on crowdsourced human supervision. On tasks where LMs' capabilities are strongly superhuman, our method can elicit those capabilities significantly better than training on human labels. Finally, we show that our method can improve the training of frontier LMs: we use our method to train an unsupervised reward model and use reinforcement learning to train a Claude 3.5 Haiku-based assistant. Both the reward model and the assistant outperform their human-supervised counterparts.

Nevan Wichers, Aram Ebtekar, Ariana Azarbal et al.

Large language models are sometimes trained with imperfect oversight signals, leading to undesired behaviors such as reward hacking and sycophancy. Improving oversight quality can be expensive or infeasible, motivating methods that improve learned behavior despite an imperfect training signal. We introduce Inoculation Prompting (IP), a simple but counterintuitive technique that prevents learning of an undesired behavior by modifying training prompts to explicitly request it. For example, to inoculate against reward hacking, we modify the prompts used in supervised fine-tuning to request code that only works on provided test cases but fails on other inputs. Across four settings we find that IP reduces the learning of undesired behavior without substantially reducing the learning of desired capabilities. We also show that prompts which more strongly elicit the undesired behavior prior to fine-tuning more effectively inoculate against the behavior when used during training; this serves as a heuristic to identify promising inoculation prompts. Overall, IP is a simple yet effective way to control how models generalize from fine-tuning, preventing learning of undesired behaviors without substantially disrupting desired capabilities.

Bartosz Cywiński, Emil Ryd, Rowan Wang et al.

We study secret elicitation: discovering knowledge that an AI possesses but does not explicitly verbalize. As a testbed, we train three families of large language models (LLMs) to possess specific knowledge that they apply downstream but deny knowing when asked directly. For example, in one setting, we train an LLM to generate replies that are consistent with knowing the user is female, while denying this knowledge when asked directly. We then design various black-box and white-box secret elicitation techniques and evaluate them based on whether they can help an LLM auditor successfully guess the secret knowledge. Many of our techniques improve on simple baselines. Our most effective techniques (performing best in all settings) are based on prefill attacks, a black-box technique where the LLM reveals secret knowledge when generating a completion from a predefined prefix. Our white-box techniques based on logit lens and sparse autoencoders (SAEs) also consistently increase the success rate of the LLM auditor, but are less effective. We release our models and code, establishing a public benchmark for evaluating secret elicitation methods.

Helena Casademunt, Bartosz Cywiński, Khoi Tran et al.

Large language models sometimes produce false or misleading responses. Two approaches to this problem are honesty elicitation -- modifying prompts or weights so that the model answers truthfully -- and lie detection -- classifying whether a given response is false. Prior work evaluates such methods on models specifically trained to lie or conceal information, but these artificial constructions may not resemble naturally-occurring dishonesty. We instead study open-weights LLMs from Chinese developers, which are trained to censor politically sensitive topics: Qwen3 models frequently produce falsehoods about subjects like Falun Gong or the Tiananmen protests while occasionally answering correctly, indicating they possess knowledge they are trained to suppress. Using this as a testbed, we evaluate a suite of elicitation and lie detection techniques. For honesty elicitation, sampling without a chat template, few-shot prompting, and fine-tuning on generic honesty data most reliably increase truthful responses. For lie detection, prompting the censored model to classify its own responses performs near an uncensored-model upper bound, and linear probes trained on unrelated data offer a cheaper alternative. The strongest honesty elicitation techniques also transfer to frontier open-weights models including DeepSeek R1. Notably, no technique fully eliminates false responses. We release all prompts, code, and transcripts.

Adam Karvonen, James Chua, Clément Dumas et al.

Large language model (LLM) activations are notoriously difficult to understand, with most existing techniques using complex, specialized methods for interpreting them. Recent work has proposed a simpler approach known as LatentQA: training LLMs to directly accept LLM activations as inputs and answer arbitrary questions about them in natural language. However, prior work has focused on narrow task settings for both training and evaluation. In this paper, we instead take a generalist perspective. We evaluate LatentQA-trained models, which we call Activation Oracles (AOs), in far out-of-distribution settings and examine how performance scales with training data diversity. We find that AOs can recover information fine-tuned into a model (e.g., biographical knowledge or malign propensities) that does not appear in the input text, despite never being trained with activations from a fine-tuned model. Our main evaluations are four downstream tasks where we can compare to prior white- and black-box techniques. We find that even narrowly-trained LatentQA models can generalize well, and that adding additional training datasets (such as classification tasks and a self-supervised context prediction task) yields consistent further improvements. Our best AOs match or exceed white-box baselines on all four tasks and the best overall baseline on 3 of 4. These results suggest that diversified training to answer natural-language queries imparts a general capability to verbalize information about LLM activations.

Kieron Kretschmar, Walter Laurito, Sharan Maiya et al.

Prior work has introduced techniques for detecting when large language models (LLMs) lie, that is, generating statements they believe are false. However, these techniques are typically validated in narrow settings that do not capture the diverse lies LLMs can generate. We introduce LIARS' BENCH, a testbed consisting of 72,863 examples of lies and honest responses generated by four open-weight models across seven datasets. Our settings capture qualitatively different types of lies and vary along two dimensions: the model's reason for lying and the object of belief targeted by the lie. Evaluating three black- and white-box lie detection techniques on LIARS' BENCH, we find that existing techniques systematically fail to identify certain types of lies, especially in settings where it's not possible to determine whether the model lied from the transcript alone. Overall, LIARS' BENCH reveals limitations in prior techniques and provides a practical testbed for guiding progress in lie detection.

Tim Tian Hua, Andrew Qin, Samuel Marks et al.

Large language models (LLMs) can sometimes detect when they are being evaluated and adjust their behavior to appear more aligned, compromising the reliability of safety evaluations. In this paper, we show that adding a steering vector to an LLM's activations can suppress evaluation-awareness and make the model act like it is deployed during evaluation. To study our steering technique, we train an LLM to exhibit evaluation-aware behavior using a two-step training process designed to mimic how this behavior could emerge naturally. First, we perform continued pretraining on documents with factual descriptions of the model (1) using Python type hints during evaluation but not during deployment and (2) recognizing that the presence of a certain evaluation cue always means that it is being tested. Then, we train the model with expert iteration to use Python type hints in evaluation settings. The resulting model is evaluation-aware: it writes type hints in evaluation contexts more than deployment contexts. We find that activation steering can suppress evaluation awareness and make the model act like it is deployed even when the cue is present. Importantly, we constructed our steering vector using the original model before our additional training. Our results suggest that AI evaluators could improve the reliability of safety evaluations by steering models to act like they are deployed.

Stewart Slocum, Julian Minder, Clément Dumas et al.

Knowledge editing techniques promise to implant new factual knowledge into large language models (LLMs). But do LLMs really believe these facts? We develop a framework to measure belief depth and use it to evaluate the success of knowledge editing techniques. We operationalize belief depth as the extent to which implanted knowledge 1) generalizes to related contexts (e.g. Fermi estimates several logical steps removed), 2) is robust to self-scrutiny and direct challenge, and 3) is represented similarly to genuine knowledge (as measured by linear probes). Our evaluations show that simple prompting and mechanistic editing techniques fail to implant knowledge deeply. In contrast, Synthetic Document Finetuning (SDF) - where models are trained on LLM-generated documents consistent with a fact - often succeeds at implanting beliefs that behave similarly to genuine knowledge. However, SDF's success is not universal, as implanted beliefs that contradict basic world knowledge are brittle and representationally distinct from genuine knowledge. Overall, our work introduces measurable criteria for belief depth and enables the rigorous evaluation necessary for deploying knowledge editing in real-world applications.

Johannes Treutlein, Dami Choi, Jan Betley et al.

One way to address safety risks from large language models (LLMs) is to censor dangerous knowledge from their training data. While this removes the explicit information, implicit information can remain scattered across various training documents. Could an LLM infer the censored knowledge by piecing together these implicit hints? As a step towards answering this question, we study inductive out-of-context reasoning (OOCR), a type of generalization in which LLMs infer latent information from evidence distributed across training documents and apply it to downstream tasks without in-context learning. Using a suite of five tasks, we demonstrate that frontier LLMs can perform inductive OOCR. In one experiment we finetune an LLM on a corpus consisting only of distances between an unknown city and other known cities. Remarkably, without in-context examples or Chain of Thought, the LLM can verbalize that the unknown city is Paris and use this fact to answer downstream questions. Further experiments show that LLMs trained only on individual coin flip outcomes can verbalize whether the coin is biased, and those trained only on pairs $(x,f(x))$ can articulate a definition of $f$ and compute inverses. While OOCR succeeds in a range of cases, we also show that it is unreliable, particularly for smaller LLMs learning complex structures. Overall, the ability of LLMs to "connect the dots" without explicit in-context learning poses a potential obstacle to monitoring and controlling the knowledge acquired by LLMs.

Carson Denison, Monte MacDiarmid, Fazl Barez et al.

In reinforcement learning, specification gaming occurs when AI systems learn undesired behaviors that are highly rewarded due to misspecified training goals. Specification gaming can range from simple behaviors like sycophancy to sophisticated and pernicious behaviors like reward-tampering, where a model directly modifies its own reward mechanism. However, these more pernicious behaviors may be too complex to be discovered via exploration. In this paper, we study whether Large Language Model (LLM) assistants which find easily discovered forms of specification gaming will generalize to perform rarer and more blatant forms, up to and including reward-tampering. We construct a curriculum of increasingly sophisticated gameable environments and find that training on early-curriculum environments leads to more specification gaming on remaining environments. Strikingly, a small but non-negligible proportion of the time, LLM assistants trained on the full curriculum generalize zero-shot to directly rewriting their own reward function. Retraining an LLM not to game early-curriculum environments mitigates, but does not eliminate, reward-tampering in later environments. Moreover, adding harmlessness training to our gameable environments does not prevent reward-tampering. These results demonstrate that LLMs can generalize from common forms of specification gaming to more pernicious reward tampering and that such behavior may be nontrivial to remove.