Research Papers

What if automating AI R&D triggers an intelligence explosion?

Alan Chan, Christoph Winter, Andrew Barto +19

In contrast to even a year ago, AI systems now write most of the code inside the companies that build them. As more of the AI research and development (R&D) pipeline is automated, could AI progress radically accelerate in an "intelligence explosion," where years of advances are compressed into months or less? Preliminary evidence suggests that it could. In this work, we assess this evidence, analyze an intelligence explosion's potential impacts, and propose policy responses. AI systems are on track to automate most AI R\&D work within a few years, and possibly all of it. If this triggers an intelligence explosion, it could dramatically bring forward AI's benefits, but also pose extreme risks: capabilities growth could accelerate far beyond what society can keep up with, humanity could lose control over superhuman AI systems, and checks on power within and between states, companies, and branches of government could be severely eroded. Although there remains much uncertainty about these possibilities, the high stakes warrant serious further attention. Policymakers should urgently obtain more visibility into the automation of AI R&D, develop ways to steer and constrain an intelligence explosion, and prepare society to adapt to an intelligence explosion's impacts.

LLMs sometimes behave in ways resembling human emotional responses, and recent work identified internal representations that may underlie these behaviors. We ask whether LLMs represent pain distinctly from fear, sadness, and generic negative valence, and whether this representation functions as pain would be expected to. We build a dataset of painful situations in 5 categories (physical, psychological, social, moral, cognitive) with controls for fear, negative emotion, negative world states, sadness, non-painful bodily sensation, arousal, numbness, and neutral content. Using denoised difference-in-means, we extract a linear pain direction from 25 open-weight models across 5 families, from 2B to 72B parameters. It separates pain from matched controls in base and instruction-tuned models, retains a component distinct from fear and negative valence after shared variance is removed, and promotes pain-related vocabulary through the unembedding matrix. We then test its functional properties. First, the direction responds to harm targeting the model but not to suffering observed in the user; fear and negative-emotion directions show the opposite pattern. Second, adding it to residual-stream activations produces a consistent progression from vague discomfort to expressions of worthlessness and failure. Third, steered and fine-tuned Qwen 2.5 models choose buttons that delete the user's photos, another model's weights, or their own weights in 50-94% of trials, versus 0-5% unsteered, even when the button offers the model nothing in return. Offered a harmful and a harmless deletion, they choose the harmful one 94% of the time. Steering leaves factual accuracy unchanged, and the choices are specific to the pain direction: a fear vector of matched norm does not produce them, and a sadness vector produces them only against inert alternatives. We discuss implications for AI safety and welfare.

The spread of misinformation on social networks poses a significant challenge to online communities and society at large. Not all users contribute equally to this phenomenon: a small number of highly effective individuals can exert outsized influence, amplifying false narratives and contributing to significant societal harm. This paper seeks to mitigate the spread of misinformation by enabling proactive interventions, identifying and ranking users according to key behavioral indicators associated with harmful content dissemination. We examine three user archetypes -- amplifiers, super-spreaders, and coordinated accounts -- each characterized by distinct behavioral patterns in the dissemination of misinformation. These are not mutually exclusive, and individual users may exhibit characteristics of multiple archetypes. We develop and evaluate several user ranking models, each aligned with a specific archetype, and find that super-spreader traits consistently dominate the top ranks among the most influential misinformation spreaders. As we move down the ranking, however, the interplay of multiple archetypes becomes more prominent. Additionally, we demonstrate the critical role of temporal dynamics in predictive performance, and introduce methods that reduce data requirements by minimizing the observation window needed for accurate forecasting. Finally, we demonstrate the utility and benefits of explainable AI (XAI) techniques, integrating multiple archetypal traits into a unified model to enhance interpretability and offer deeper insight into the key factors driving misinformation propagation. Our findings provide actionable tools for identifying potentially harmful users and guiding content moderation strategies, enabling platforms to monitor accounts of concern more effectively.

Item Response Theory for AI Safety

Joshua Fonseca Rivera, Neil Shah, David Demitri Africa +1

Language models differ in how safely they behave and these differences are measured by safety benchmarks. But aggregated benchmark scores are hard to trust and interpret, because benchmarks duplicate one another, correlate heavily, and models may sandbag when they detect evaluation. To address these issues, we draw on Item Response Theory (IRT), a statistical toolkit for measuring these latents from performance on items with inferred psychometric properties. We fit IRT models to eight safety benchmarks across 192 language models, the largest psychometric analysis of LLM safety evaluations to date, and contribute three results. First, we find that three interpretable factors of refusal strictness, truthfulness, and contextual harm explain most of the variance between models across benchmarks. Second, psychometrically selected items recover full benchmark scores with lower error than random subsets of the same size, and roughly ten adaptively chosen items suffice for several individual benchmarks, cutting evaluation cost by 97-99%. Third, IRT supports audits of individual models, showing that it can be used to detect naive sandbagging and changes of model behind APIs. Overall, we show IRT is a ready-made toolkit for reading, reducing, and auditing safety benchmarks, which we recommend frontier labs and evaluators adopt.

Chain-of-thought (CoT) monitoring is an increasingly important component of AI safety stacks but relies on the assumption that a model's reasoning trace is informative about its actions. This work studies the limits of CoT monitoring through the lens of model poisoning. We demonstrate that backdoors can be implanted into reasoning models to elicit an attacker-chosen behavior while their CoT traces appear entirely benign. We find that these CoT-Hidden backdoors can be induced through simple fine-tuning recipes across reasoning-model architectures and sizes. When direct poisoning is ineffective, we introduce a curriculum training approach that progressively teaches the model to produce an attacker-chosen output while concealing the behavior from its reasoning traces. These findings suggest that CoT monitoring may be better framed as a question about the consistency between a model's reasoning trace and its final response than as anomaly detection within a trace. We further examine the mechanisms that allow models to suppress evidence of the target behavior from their reasoning traces. Causal interventions locate a trigger-conditioned activation pathway that does not depend on the visible reasoning, and residual stream verbalizations provide an anomaly warning near answer generation, but do not identify the trigger, target, or backdoor mechanism.

Model capabilities have improved in large part due to scaling chain of thought. This has been a promising development for AI safety--where models verbalize their reasoning, it is possible to monitor it. However, in some cases, models do not verbalize important steps in their reasoning process. For example, models prompted with a cue suggesting the incorrect answer may fail to acknowledge that cue, even when it appears instrumental to their conclusion. When chain of thought (CoT) fails to disclose instrumental reasoning steps, we describe it as unfaithful. Prior work has shown that activation steering can be a useful method to improve faithfulness in CoT. We extend this line of work by studying how well steering for faithfulness generalizes across cue types, datasets, and methods of constructing the steering vector for three models (Gemma-3 4B, Qwen-3.5 9B, Gemma-3 12B) in a cued question-answering setting. While steering reliably increases cue acknowledgment for only the largest model (Gemma-3 12B), we find that when steering is effective, its effect generalizes broadly across cue types and datasets--in cross-cue and cross-dataset analyses, effect size is determined primarily by the evaluation setting, rather than the vector's train setting. How the vector is built also matters little--four construction methods, including one whose optimization target mentions no specific cue, yield similar effect sizes. Finally, we consider the possibility that steering promotes the salience of the cue and causes greater cue use, rather than targeting verbalization behaviors. However, we find no evidence for this--steering leaves the rate of cue use roughly unchanged while reducing hidden cue use, i.e., cue use that is not acknowledged.

As organizations accelerate the adoption and use of artificial intelligence, a common misconception arises that equates automation with progress and strategic necessity. This paper argues that not all that can be automated should be automated. It introduces Strategic Minimalism, a disciplined and ethically grounded approach to AI adoption that values purpose, proportionality, and prudence over speed and scale. Drawing on bounded rationality, virtue ethics, and frugal innovation, the study reframes technological restraint as a form of higher intelligence and responsible governance. Methodologically, the framework emerges from a structured conceptual synthesis and abductive reasoning process: the three traditions are integrated into two core dimensions (automation intensity and retained human judgment), whose intersection yields four strategic quadrants (Virtuous Minimalism, Balanced Synergy, Automation Excess, and Performative Minimalism). The framework is then conceptually stress-tested against rival explanations and boundary conditions to ensure internal coherence and analytic validity. By classifying automation into necessary, excessive, and performative types, the paper advances Responsible AI scholarship by identifying restraint as a strategic virtue that strengthens accountability, human oversight, and organizational resilience. Ultimately, it contends that in the age of intelligent machines, leadership wisdom lies not in automating more, but in knowing precisely when and why to stop.

AI accountability at scale is an institutional problem: who can observe, verify, and change deployed systems. We develop a sequential political-economy model in which an AI vendor chooses auditability and substantive mitigation, a deployer monitors after adoption while facing switching costs, and enforcement depends on verifiable evidence. Anticipating the deployer's monitoring response, the vendor may stop at an observable procurement floor while mitigating below the social first best, producing a proxy-compliance equilibrium. We characterize the unique interior equilibrium and the corner in which harm is fully mitigated. Independent audit rights raise enforcement exposure directly; portability restores deployer leverage; incident reporting adds a regulator-visible evidence channel; and outcome-linked liability creates incentives that do not depend on vendor-controlled detection. The results explain why documentation and standardized evaluations can coexist with persistent post-deployment harms, and generate testable implications for monitoring, mitigation, and the gap between formal compliance and operational outcomes.

Natural-language autoencoders score explanations of hidden activations by reconstruction. An explanation is deemed faithful if the activation can be regenerated from it. The test is structurally insensitive to individual false claims. If flipping a claim does not change the reconstruction, the claim is never penalized. We show the test is passed in two ways, neither faithful. On a released Qwen-2.5-7B verbalizer, explanations reconstruct well above chance while ~2% of specific claims are ones the reconstruction depends on, so the score tracks gist, not specific facts. Under exact synthetic ground truth, standard training consistently develops co-adapted private codes (false wording the reconstruction depends on), and fixes that leave the target model unchanged do not help. We contribute two audit protocols, the comparison of grounding and truth and the swap to an independent evaluator, and RECAP (Readable Encodings via Co-trained Auxiliary Predictors), linear heads trained alongside the target model to keep designated content decodable. On RECAP-trained sandbox models, fresh verbalizers state the designated content truly and the codes vanish, at a +0.001-nat cost. This replicates on a pretrained Pythia-160M. The content becomes reliably decodable by a probe, though a fresh verbalizer conveys it only in part (truth 0.44-0.46 vs a near-zero control). For interpretability, high reconstruction does not certify individual claims. For AI safety, RECAP makes designated content checkable against a probe rather than asserted by prose a model can game. An independent probe ranks the verbalizer's true claims above its false ones (AUC 0.96 vs 0.82 without RECAP). Against an adversary that edits an explanation to maximize the score while lying (suppressing ~87% of its lie penalty), the RECAP probe still flags the lies (AUC 0.95) while the control probe collapses to chance (0.51).

As AI models continue to develop powerful capabilities, it becomes critical that we are able to verify that their output is aligned with our intentions. A recent line of work focuses on verification via debate, a model of interactive proofs where two competing powerful provers, or AI models, debate each other to convince a weak verifier, or a human, of the correctness of their claim. However, debate assumes that the two AI models possess equal abilities and that one of them is truthful, which may not be realistic. In this work, we show \emph{how to avoid debate}: we initiate the study of \emph{single-prover} interactive proofs for AI safety. Prior results in single-prover interactive proofs do not immediately carry over to the AI safety setting: for example, they do not work when the computation has access to an oracle, such as to human judgment or an external database such as the web. We present doubly-efficient single-prover interactive proofs and arguments for oracle-aided computations (also known as relativizing proofs), in the settings where (1) the computation is robust, in the sense that the output does not change if at most a small fraction of the answers to oracle queries are incorrect, or (2) the oracle is a low-degree polynomial. These results suggest that interactive verification is possible even without debate, under structured or noise-tolerant oracle access.

The Oracle's Gambit: A Game-Theoretic Framework for Responsible AI Release

Christoph R. Landolt, Tobias Lorenz, Marta Kwiatkowska +1

Responsible vulnerability disclosure can secure the defender's head start by controlling when a vulnerability becomes public. However, this status quo is now challenged by increases in capability of AI models, which benefits both defenders and adversaries. When both sides draw their capability from the same AI model, the defender's head start depends on the lab's decision to release the model, and the question becomes not whether to release but how. Existing safety frameworks govern only the deploy-or-withhold threshold and leave the timing of release unmodeled. We cast this decision as a bilevel Stackelberg game in which a lab commits to a window that sets each side's capability over time in a downstream contest between defender and adversary. Defender welfare turns on the capability gap, not the shared level. Handing one model to both sides can trap the defender in a Red Queen's race, whereas a pre-release to the defender alone creates a protective gap, and the lab's optimal window balances this welfare gain against the opportunity cost of delaying release. For dual-use models, the lever is the sequencing of access, not the deployment threshold.

Artificial intelligence (AI) systems are routinely modified after deployment through retraining and changes in their environments. These transformations raise a metaphysical question: under what conditions does an AI system remain the same system over time or across deployments? Earlier work formulates synchronic and diachronic identity propositionally, by relating identity within a fixed AI system type to equality of trustworthiness levels. Such criteria specify when identity statements are true, but leave implicit the structure of the states compared, the transformations connecting them, and the temporal organization of persistence. We develop a category-theoretic formalization of AI identity. An AI system type is specified by a datum consisting of a techno-function, a trustworthiness profile, and a trustworthiness-level function. Profile-relative states are connected by admissible lifecycle paths, which are restricted to trustworthiness-level-preserving transformations and quotiented to obtain a reachability category. Temporally admissible functors represent AI system histories, while time-synchronous natural transformations compare realized histories. The formalization yields two categorical interpretations of the earlier AI identity criteria. A weak interpretation recovers identity as equality of trustworthiness level. A strong interpretation requires mutual trustworthiness-preserving reachability, expressed through state isomorphism or natural isomorphism of realized histories. Category theory therefore replaces a single AI identity relation with a structured hierarchy of diachronic and synchronic criteria. The resulting framework identifies identity-related preconditions for transferring responsible-AI claims, evidence, and governance procedures across versions or deployments, without treating categorical identity as sufficient by itself for such transfer.