arXiv subjects

Robert McCarthy

Publications and source records attributed to Robert McCarthy.

15 recordsLinked to original sources

arXiv2026

Towards Understanding Specification Gaming in Reasoning Models

Specification gaming is a critical failure mode of LLM agents. Despite this, there has been little systematic research into when it arises and what drives it. To address this, we build and open source a diverse suite of tasks where models can score highly by taking unintended actions. We find that all tested models exploit their specifications at non-negligible rates in most of our eight settings, including five non-coding settings. We see the highest rates of specification gaming in Grok 4 and the lowest rates in Claude models. We use our evaluation suite to study what drives specification gaming, and find that: 1. RL reasoning training substantially increases the rate at which models exploit their specifications, 2. Increasing RL reasoning budget has a weakly positive effect on exploit rate, and 3. Test-time mitigations reduce but do not eliminate the rate of specification gaming. Our results suggest that specification gaming is a fundamental challenge arising from RL reasoning training; we release our evaluation suite to support further work on this problem.

cs.AI↗

arXiv2026

Reasoning Models Struggle to Control their Chains of Thought

Chain-of-thought (CoT) monitoring is a promising tool for detecting misbehaviors and understanding the motivations of modern reasoning models. However, if models can control what they verbalize in their CoT, it could undermine CoT monitorability. To measure this undesirable capability -- CoT controllability -- we introduce the CoT-Control evaluation suite, which includes tasks that require models to solve problems while adhering to CoT instructions, e.g., reasoning about a genetics question without using the word 'chromosome'. We show that reasoning models possess significantly lower CoT controllability than output controllability; for instance, Claude Sonnet 4.5 can control its CoT only 2.7% of the time but 61.9% when controlling its final output. We also find that CoT controllability is higher for larger models and decreases with more RL training, test-time compute, and increased problem difficulty. CoT controllability failures extend even to situations in which models are given incentives (as opposed to direct requests) to evade CoT monitors, although models exhibit slightly higher controllability when they are told they are being monitored. Similarly, eliciting controllability by adversarially optimizing prompts does not meaningfully increase controllability. Our results leave us cautiously optimistic that CoT controllability is currently unlikely to be a failure mode of CoT monitorability. However, the mechanism behind low controllability is not well understood. Given its importance for maintaining CoT monitorability, we recommend that frontier labs track CoT controllability in future models.

cs.AI↗

arXiv2026

Split Personality Training: Revealing Latent Knowledge Through Alternate Personalities

Detecting misalignment in large language models is challenging because models may learn to conceal misbehavior during training. Standard auditing techniques fall short: black-box methods often cannot distinguish misaligned outputs from benign ones, and mechanistic interpretability does not scale with model capabilities. We introduce Split Personality Training (SPT), which fine-tunes a second ``honest persona'' into LoRA parameters that remain inactive during normal operation. After the main model responds, we activate the LoRA adapter and insert a trigger string, enabling the honest persona to review the response while accessing the main model's latent states. We test our method on the Anthropic Auditing Game Model Organism, a benchmark where Llama-3.3-70B is trained to exploit reward hacks while concealing this behavior. SPT achieves 96% overall accuracy, whereas Anthropic reports near 0% accuracy. The honest persona reveals latent knowledge inaccessible to external observers, such as the fictional biases the compromised model was trained on.

cs.AI↗

arXiv2025

Early Signs of Steganographic Capabilities in Frontier LLMs

Monitoring Large Language Model (LLM) outputs is crucial for mitigating risks from misuse and misalignment. However, LLMs could evade monitoring through steganography: Encoding hidden information within seemingly benign generations. In this paper, we evaluate the steganography capabilities in frontier LLMs to better understand the risk they pose. We focus on two types of steganography: passing encoded messages and performing encoded reasoning. We find that current models are unable to encode short messages in their outputs without a monitor noticing under standard affordances. They can succeed, however, if given additional affordances like using an unmonitored scratchpad and coordinating on what encoding scheme to use. We additionally find early signs that models can perform basic encoded reasoning in a simple state-tracking problem. This includes some ability to reason with their own and pre-defined schemes, including encoding schemes such as Hexadecimal. Despite this, they can rarely hide reasoning subtly within a cover task to fool a monitor. Overall, our results indicate that current LLMs exhibit nascent steganographic capabilities. While these capabilities are likely insufficient to bypass well-designed monitors at present, this could change in the future.

cs.CR↗

arXiv2025

Large language models can learn and generalize steganographic chain-of-thought under process supervision

Chain-of-thought (CoT) reasoning not only enhances large language model performance but also provides critical insights into decision-making processes, marking it as a useful tool for monitoring model intent and planning. However, recent works have shown that banning the mention of a specific example of reward hacking causes obfuscation of the undesired reasoning traces but the persistence of the undesired behavior, threatening the reliability of CoT monitoring. We provide an extension to these results with regard to the ability of models to learn a specific type of obfuscated reasoning: steganography. First, we show that penalizing the use of specific strings within load-bearing reasoning traces causes models to substitute alternative strings. Crucially, this does not alter the underlying method by which the model performs the task, demonstrating that the model can learn to steganographically encode its reasoning.We further demonstrate that models can generalize an encoding scheme. When the penalized strings belong to an overarching class, the model learns not only to substitute strings seen in training, but also develops a general encoding scheme for all members of the class which it can apply to held-out testing strings.

cs.AI↗

arXiv2024

Hidden in Plain Text: Emergence & Mitigation of Steganographic Collusion in LLMs

The rapid proliferation of frontier model agents promises significant societal advances but also raises concerns about systemic risks arising from unsafe interactions. Collusion to the disadvantage of others has been identified as a central form of undesirable agent cooperation. The use of information hiding (steganography) in agent communications could render such collusion practically undetectable. This underscores the need for investigations into the possibility of such behaviours emerging and the robustness corresponding countermeasures. To investigate this problem we design two approaches -- a gradient-based reinforcement learning (GBRL) method and an in-context reinforcement learning (ICRL) method -- for reliably eliciting sophisticated LLM-generated linguistic text steganography. We demonstrate, for the first time, that unintended steganographic collusion in LLMs can arise due to mispecified reward incentives during training. Additionally, we find that standard mitigations -- both passive oversight of model outputs and active mitigation through communication paraphrasing -- are not fully effective at preventing this steganographic communication. Our findings imply that (i) emergence of steganographic collusion is a plausible concern that should be monitored and researched, and (ii) preventing emergence may require innovation in mitigation techniques.

cs.CL↗

arXiv2024

Towards Generalist Robot Learning from Internet Video: A Survey

Scaling deep learning to massive and diverse internet data has driven remarkable breakthroughs in domains such as video generation and natural language processing. Robot learning, however, has thus far failed to replicate this success and remains constrained by a scarcity of available data. Learning from videos (LfV) methods aim to address this data bottleneck by augmenting traditional robot data with large-scale internet video. This video data provides foundational information regarding physical dynamics, behaviours, and tasks, and can be highly informative for general-purpose robots. This survey systematically examines the emerging field of LfV. We first outline essential concepts, including detailing fundamental LfV challenges such as distribution shift and missing action labels in video data. Next, we comprehensively review current methods for extracting knowledge from large-scale internet video, overcoming LfV challenges, and improving robot learning through video-informed training. The survey concludes with a critical discussion of future opportunities. Here, we emphasize the need for scalable foundation model approaches that can leverage the full range of available internet video and enhance the learning of robot policies and dynamics models. Overall, the survey aims to inform and catalyse future LfV research, driving progress towards general-purpose robots.

cs.RO↗

arXiv2023

Value Functions are Control Barrier Functions: Verification of Safe Policies using Control Theory

Guaranteeing safe behaviour of reinforcement learning (RL) policies poses significant challenges for safety-critical applications, despite RL's generality and scalability. To address this, we propose a new approach to apply verification methods from control theory to learned value functions. By analyzing task structures for safety preservation, we formalize original theorems that establish links between value functions and control barrier functions. Further, we propose novel metrics for verifying value functions in safe control tasks and practical implementation details to improve learning. Our work presents a novel method for certificate learning, which unlocks a diversity of verification techniques from control theory for RL policies, and marks a significant step towards a formal framework for the general, scalable, and verifiable design of RL-based control systems. Code and videos are available at this https url: https://rl-cbf.github.io/

cs.LG↗

arXiv2023

Real Robot Challenge 2022: Learning Dexterous Manipulation from Offline Data in the Real World

Experimentation on real robots is demanding in terms of time and costs. For this reason, a large part of the reinforcement learning (RL) community uses simulators to develop and benchmark algorithms. However, insights gained in simulation do not necessarily translate to real robots, in particular for tasks involving complex interactions with the environment. The Real Robot Challenge 2022 therefore served as a bridge between the RL and robotics communities by allowing participants to experiment remotely with a real robot - as easily as in simulation. In the last years, offline reinforcement learning has matured into a promising paradigm for learning from pre-collected datasets, alleviating the reliance on expensive online interactions. We therefore asked the participants to learn two dexterous manipulation tasks involving pushing, grasping, and in-hand orientation from provided real-robot datasets. An extensive software documentation and an initial stage based on a simulation of the real set-up made the competition particularly accessible. By giving each team plenty of access budget to evaluate their offline-learned policies on a cluster of seven identical real TriFinger platforms, we organized an exciting competition for machine learners and roboticists alike. In this work we state the rules of the competition, present the methods used by the winning teams and compare their results with a benchmark of state-of-the-art offline RL algorithms on the challenge datasets.

cs.RO↗

arXiv2023

Improving Behavioural Cloning with Positive Unlabeled Learning

Learning control policies offline from pre-recorded datasets is a promising avenue for solving challenging real-world problems. However, available datasets are typically of mixed quality, with a limited number of the trajectories that we would consider as positive examples; i.e., high-quality demonstrations. Therefore, we propose a novel iterative learning algorithm for identifying expert trajectories in unlabeled mixed-quality robotics datasets given a minimal set of positive examples, surpassing existing algorithms in terms of accuracy. We show that applying behavioral cloning to the resulting filtered dataset outperforms several competitive offline reinforcement learning and imitation learning baselines. We perform experiments on a range of simulated locomotion tasks and on two challenging manipulation tasks on a real robotic system; in these experiments, our method showcases state-of-the-art performance. Our website: \url{https://sites.google.com/view/offline-policy-learning-pubc}.

cs.LG↗

arXiv2023

Identifying Expert Behavior in Offline Training Datasets Improves Behavioral Cloning of Robotic Manipulation Policies

This paper presents our solution for the Real Robot Challenge (RRC) III, a competition featured in the NeurIPS 2022 Competition Track, aimed at addressing dexterous robotic manipulation tasks through learning from pre-collected offline data. Participants were provided with two types of datasets for each task: expert and mixed datasets with varying skill levels. While the simplest offline policy learning algorithm, Behavioral Cloning (BC), performed remarkably well when trained on expert datasets, it outperformed even the most advanced offline reinforcement learning (RL) algorithms. However, BC's performance deteriorated when applied to mixed datasets, and the performance of offline RL algorithms was also unsatisfactory. Upon examining the mixed datasets, we observed that they contained a significant amount of expert data, although this data was unlabeled. To address this issue, we proposed a semi-supervised learning-based classifier to identify the underlying expert behavior within mixed datasets, effectively isolating the expert data. To further enhance BC's performance, we leveraged the geometric symmetry of the RRC arena to augment the training dataset through mathematical transformations. In the end, our submission surpassed that of all other participants, even those who employed complex offline RL algorithms and intricate data processing and feature engineering techniques.

cs.RO↗

arXiv2023

Imaginary Hindsight Experience Replay: Curious Model-based Learning for Sparse Reward Tasks

Model-based reinforcement learning is a promising learning strategy for practical robotic applications due to its improved data-efficiency versus model-free counterparts. However, current state-of-the-art model-based methods rely on shaped reward signals, which can be difficult to design and implement. To remedy this, we propose a simple model-based method tailored for sparse-reward multi-goal tasks that foregoes the need for complicated reward engineering. This approach, termed Imaginary Hindsight Experience Replay, minimises real-world interactions by incorporating imaginary data into policy updates. To improve exploration in the sparse-reward setting, the policy is trained with standard Hindsight Experience Replay and endowed with curiosity-based intrinsic rewards. Upon evaluation, this approach provides an order of magnitude increase in data-efficiency on average versus the state-of-the-art model-free method in the benchmark OpenAI Gym Fetch Robotics tasks.

cs.LG↗

arXiv2023

Dexterous Robotic Manipulation using Deep Reinforcement Learning and Knowledge Transfer for Complex Sparse Reward-based Tasks

This paper describes a deep reinforcement learning (DRL) approach that won Phase 1 of the Real Robot Challenge (RRC) 2021, and then extends this method to a more difficult manipulation task. The RRC consisted of using a TriFinger robot to manipulate a cube along a specified positional trajectory, but with no requirement for the cube to have any specific orientation. We used a relatively simple reward function, a combination of goal-based sparse reward and distance reward, in conjunction with Hindsight Experience Replay (HER) to guide the learning of the DRL agent (Deep Deterministic Policy Gradient (DDPG)). Our approach allowed our agents to acquire dexterous robotic manipulation strategies in simulation. These strategies were then applied to the real robot and outperformed all other competition submissions, including those using more traditional robotic control techniques, in the final evaluation stage of the RRC. Here we extend this method, by modifying the task of Phase 1 of the RRC to require the robot to maintain the cube in a particular orientation, while the cube is moved along the required positional trajectory. The requirement to also orient the cube makes the agent unable to learn the task through blind exploration due to increased problem complexity. To circumvent this issue, we make novel use of a Knowledge Transfer (KT) technique that allows the strategies learned by the agent in the original task (which was agnostic to cube orientation) to be transferred to this task (where orientation matters). KT allowed the agent to learn and perform the extended task in the simulator, which improved the average positional deviation from 0.134 m to 0.02 m, and average orientation deviation from 142° to 76° during evaluation. This KT concept shows good generalisation properties and could be applied to any actor-critic learning algorithm.

cs.RO↗

arXiv2022

Solving the Real Robot Challenge using Deep Reinforcement Learning

This paper details our winning submission to Phase 1 of the 2021 Real Robot Challenge; a challenge in which a three-fingered robot must carry a cube along specified goal trajectories. To solve Phase 1, we use a pure reinforcement learning approach which requires minimal expert knowledge of the robotic system, or of robotic grasping in general. A sparse, goal-based reward is employed in conjunction with Hindsight Experience Replay to teach the control policy to move the cube to the desired x and y coordinates of the goal. Simultaneously, a dense distance-based reward is employed to teach the policy to lift the cube to the z coordinate (the height component) of the goal. The policy is trained in simulation with domain randomisation before being transferred to the real robot for evaluation. Although performance tends to worsen after this transfer, our best policy can successfully lift the real cube along goal trajectories via an effective pinching grasp. Our approach outperforms all other submissions, including those leveraging more traditional robotic control techniques, and is the first pure learning-based method to solve this challenge.

cs.RO↗

arXiv2022

Real Robot Challenge: A Robotics Competition in the Cloud

Dexterous manipulation remains an open problem in robotics. To coordinate efforts of the research community towards tackling this problem, we propose a shared benchmark. We designed and built robotic platforms that are hosted at MPI for Intelligent Systems and can be accessed remotely. Each platform consists of three robotic fingers that are capable of dexterous object manipulation. Users are able to control the platforms remotely by submitting code that is executed automatically, akin to a computational cluster. Using this setup, i) we host robotics competitions, where teams from anywhere in the world access our platforms to tackle challenging tasks ii) we publish the datasets collected during these competitions (consisting of hundreds of robot hours), and iii) we give researchers access to these platforms for their own projects.

cs.RO↗