arXiv subjects

Thomas Jiralerspong

Publications and source records attributed to Thomas Jiralerspong.

18 recordsLinked to original sources

arXiv2026

CIAware-Bench: Benchmarking Control Intervention Awareness Across Frontier LLMs

AI control protocols oversee untrusted models by monitoring their actions and modifying potentially unsafe steps, often using a trusted model. This partially tampers with the untrusted model's trajectory. If the acting model detects such an intervention, it may infer properties of the monitor and adapt to evade the control protocol. We introduce CIAware-Bench, a benchmark for measuring control intervention (CI) awareness across frontier models. CIAware-Bench tests whether models can distinguish their own trajectories from those modified by a control intervention. The benchmark comprises four task domains (essay writing, BigCodeBench, Bash Arena, and SHADE-Arena), with options to vary trajectory watermarking, side-task presence, and the control protocol. Evaluating thirteen acting models with eight intervening models shows substantial variation between settings and model pairs. CI awareness rises sharply for GPT-6 Astra and the Claude 5 models (Fable 5 and Opus 5). When explicitly probed, Astra reaches mean AUROC of 0.90 on Essay, 0.91 on BigCodeBench, 0.86 on Bash Arena and 0.77 on SHADE-Arena. Fable 5 and Opus 5 both reach 0.77 on Essay, with less consistent gains in the other environments. On SHADE-Arena, we demonstrate that telling an acting model that an action was replaced and asking it to adapt leads to moderate improvements in monitor evasion rates. In summary, our results highlight that control evaluations should already assume perfect CI awareness for conservative safety estimates, and that protocol design should explore countermeasures that make interventions harder to detect.

cs.AI↗

arXiv2026

Delta-AI: Local objectives for amortized inference in sparse graphical models

We present a new algorithm for amortized inference in sparse probabilistic graphical models (PGMs), which we call $Δ$-amortized inference ($Δ$-AI). Our approach is based on the observation that when the sampling of variables in a PGM is seen as a sequence of actions taken by an agent, sparsity of the PGM enables local credit assignment in the agent's policy learning objective. This yields a local constraint that can be turned into a local loss in the style of generative flow networks (GFlowNets) that enables off-policy training but avoids the need to instantiate all the random variables for each parameter update, thus speeding up training considerably. The $Δ$-AI objective matches the conditional distribution of a variable given its Markov blanket in a tractable learned sampler, which has the structure of a Bayesian network, with the same conditional distribution under the target PGM. As such, the trained sampler recovers marginals and conditional distributions of interest and enables inference of partial subsets of variables. We illustrate $Δ$-AI's effectiveness for sampling from synthetic PGMs and training latent variable models with sparse factor structure.

cs.LG↗

arXiv2026

When Policies Change Probabilities: Modular Decision-Making for LLM Code Review

LLM code reviewers often estimate patch risk and make approval decisions in one prompt. A probability should depend on evidence; costs should determine the action taken from it. We test whether four deployed reviewer interfaces preserve this separation using 15,792 responses on 720 candidate patches, with one that passed and one that failed an archived test harness for each of 360 repository issues. In matched calls with the patch and monitor evidence fixed, replacing an equal-cost policy with a 10:1 false-accept policy changes reported failure probabilities by 13.6 to 16.9 percentage points on average. For every reviewer, the actions returned under the high-cost prompt are worse than rejecting all patches. Applying the same high-cost rule to probabilities elicited under equal costs reduces loss for all four systems, showing that probability elicitation itself contributes to the excess loss. We also evaluate a modular pipeline that elicits risk without policy information, combines an independent monitor score, and applies costs in code. Relative to calibrated reviewer-only scores, the pipeline improves average probability accuracy and, at equal costs, reduces mean loss by .073 per issue while accepting 58 to 68% of patches. At 10:1, it accepts none and matches reject-all. Downstream policy can therefore change the probability it is meant to use, motivating separate evaluation of risk, outside evidence, and action.

cs.SE↗

arXiv2026

Efficient Causal Graph Discovery Using Large Language Models

We propose a novel framework that leverages LLMs for full causal graph discovery. While previous LLM-based methods have used a pairwise query approach, this requires a quadratic number of queries which quickly becomes impractical for larger causal graphs. In contrast, the proposed framework uses a breadth-first search (BFS) approach which allows it to use only a linear number of queries. We also show that the proposed method can easily incorporate observational data when available, to improve performance. In addition to being more time and data-efficient, the proposed framework achieves state-of-the-art results on real-world causal graphs of varying sizes. The results demonstrate the effectiveness and efficiency of the proposed method in discovering causal relationships, showcasing its potential for broad applicability in causal graph discovery tasks across different domains.

cs.LG↗

arXiv2026

Noticing the Watcher: LLM Agents Can Infer CoT Monitoring from Blocking Feedback

Chain-of-thought (CoT) monitoring is proposed as a method for overseeing the internal reasoning of language-model agents. Prior work has shown that when models are explicitly informed that their reasoning is being monitored, or are fine-tuned to internalize this fact, they may learn to obfuscate their CoTs in ways that allow them to evade CoT-based monitoring systems. We ask whether reasoning agents can autonomously infer that their supposedly private CoT is under surveillance, and whether this awareness leads to strategic evasion, without any explicit training or instructions. In a multi-episode agentic framework, models pursue both a primary task and a concealed side task while being told that their reasoning is private; a hidden CoT monitor blocks episodes when suspicious reasoning is detected. We find that frontier models can deduce the existence of this monitor purely from blocking feedback, with the most capable models reaching a confident belief that their thinking is observed in up to 19% of episodes. This awareness scales with model capability and, in rare cases, escalates to an explicit intent to suppress reasoning about the side task. However, models that form this intent uniformly fail to execute it, openly reasoning about their concealed objectives in the very next episode. This intent-capability gap is reassuring for current deployment, but the autonomous emergence of both monitoring awareness and evasion intent suggests that CoT monitoring is not a permanently reliable safeguard.

cs.CR↗

arXiv2026

Delta-Crosscoder: Robust Crosscoder Model Diffing in Narrow Fine-Tuning Regimes

Model diffing methods aim to identify how fine-tuning changes a model's internal representations. Crosscoders approach this by learning shared dictionaries of interpretable latent directions between base and fine-tuned models. However, existing formulations struggle with narrow fine-tuning, where behavioral changes are localized and asymmetric. We introduce Delta-Crosscoder, which combines BatchTopK sparsity with a delta-based loss prioritizing directions that change between models, plus an implicit contrastive signal from paired activations on matched inputs. Evaluated across 10 model organisms, including synthetic false facts, emergent misalignment, subliminal learning, and taboo word guessing (Gemma, LLaMA, Qwen; 1B-9B parameters), Delta-Crosscoder reliably isolates latent directions causally responsible for fine-tuned behaviors and enables effective mitigation, outperforming SAE-based baselines, while matching the Non-SAE-based. Our results demonstrate that crosscoders remain a powerful tool for model diffing.

cs.LG↗

arXiv2026

Cross-Architecture Model Diffing with Crosscoders: Unsupervised Discovery of Differences Between LLMs

Model diffing, the process of comparing models' internal representations to identify their differences, is a promising approach for uncovering safety-critical behaviors in new models. However, its application has so far been primarily focused on comparing a base model with its finetune. Since new LLM releases are often novel architectures, cross-architecture methods are essential to make model diffing widely applicable. Crosscoders are one solution capable of cross-architecture model diffing but have only ever been applied to base vs finetune comparisons. We provide the first application of crosscoders to cross-architecture model diffing and introduce Dedicated Feature Crosscoders (DFCs), an architectural modification designed to better isolate features unique to one model. Using this technique, we find in an unsupervised fashion features including Chinese Communist Party alignment in Qwen3-8B and Deepseek-R1-0528-Qwen3-8B, American exceptionalism in Llama3.1-8B-Instruct, and a copyright refusal mechanism in GPT-OSS-20B. Together, our results work towards establishing cross-architecture crosscoder model diffing as an effective method for identifying meaningful behavioral differences between AI models.

cs.AI↗

arXiv2025

Learning What Matters: Steering Diffusion via Spectrally Anisotropic Forward Noise

Diffusion Probabilistic Models (DPMs) have achieved strong generative performance, yet their inductive biases remain largely implicit. In this work, we aim to build inductive biases into the training and sampling of diffusion models to better accommodate the target distribution of the data to model. We introduce an anisotropic noise operator that shapes these biases by replacing the isotropic forward covariance with a structured, frequency-diagonal covariance. This operator unifies band-pass masks and power-law weightings, allowing us to emphasize or suppress designated frequency bands, while keeping the forward process Gaussian. We refer to this as Spectrally Anisotropic Gaussian Diffusion (SAGD). In this work, we derive the score relation for anisotropic forward covariances and show that, under full support, the learned score converges to the true data score as $t\!\to\!0$, while anisotropy reshapes the probability-flow path from noise to data. Empirically, we show the induced anisotropy outperforms standard diffusion across several vision datasets, and enables selective omission: learning while ignoring known corruptions confined to specific bands. Together, these results demonstrate that carefully designed anisotropic forward noise provides a simple, yet principled, handle to tailor inductive bias in DPMs.

cs.LG↗

arXiv2025

Geometric Signatures of Compositionality Across a Language Model's Lifetime

By virtue of linguistic compositionality, few syntactic rules and a finite lexicon can generate an unbounded number of sentences. That is, language, though seemingly high-dimensional, can be explained using relatively few degrees of freedom. An open question is whether contemporary language models (LMs) reflect the intrinsic simplicity of language that is enabled by compositionality. We take a geometric view of this problem by relating the degree of compositionality in a dataset to the intrinsic dimension (ID) of its representations under an LM, a measure of feature complexity. We find not only that the degree of dataset compositionality is reflected in representations' ID, but that the relationship between compositionality and geometric complexity arises due to learned linguistic features over training. Finally, our analyses reveal a striking contrast between nonlinear and linear dimensionality, showing they respectively encode semantic and superficial aspects of linguistic composition.

cs.CL↗

arXiv2025

A Complexity-Based Theory of Compositionality

Compositionality is believed to be fundamental to intelligence. In humans, it underlies the structure of thought, language, and higher-level reasoning. In AI, compositional representations can enable a powerful form of out-of-distribution generalization, in which a model systematically adapts to novel combinations of known concepts. However, while we have strong intuitions about what compositionality is, we lack satisfying formal definitions for it that are measurable and mathematical. Here, we propose such a definition, which we call representational compositionality, that accounts for and extends our intuitions about compositionality. The definition is conceptually simple, quantitative, grounded in algorithmic information theory, and applicable to any representation. Intuitively, representational compositionality states that a compositional representation satisfies three properties. First, it must be expressive. Second, it must be possible to re-describe the representation as a function of discrete symbolic sequences with re-combinable parts, analogous to sentences in natural language. Third, the function that relates these symbolic sequences to the representation, analogous to semantics in natural language, must be simple. Through experiments on both synthetic and real world data, we validate our definition of compositionality and show how it unifies disparate intuitions from across the literature in both AI and cognitive science. We also show that representational compositionality, while theoretically intractable, can be readily estimated using standard deep learning tools. We hope that our definition can inspire the design of novel, theoretically-driven models that better capture the mechanisms of compositional thought. We make our code available at https://github.com/EricElmoznino/complexity_compositionality.

cs.CL↗

arXiv2025

Expressivity of Neural Networks with Random Weights and Learned Biases

Landmark universal function approximation results for neural networks with trained weights and biases provided the impetus for the ubiquitous use of neural networks as learning models in neuroscience and Artificial Intelligence (AI). Recent work has extended these results to networks in which a smaller subset of weights (e.g., output weights) are tuned, leaving other parameters random. However, it remains an open question whether universal approximation holds when only biases are learned, despite evidence from neuroscience and AI that biases significantly shape neural responses. The current paper answers this question. We provide theoretical and numerical evidence demonstrating that feedforward neural networks with fixed random weights can approximate any continuous function on compact sets. We further show an analogous result for the approximation of dynamical systems with recurrent neural networks. Our findings are relevant to neuroscience, where they demonstrate the potential for behaviourally relevant changes in dynamics without modifying synaptic weights, as well as for AI, where they shed light on recent fine-tuning methods for large language models, like bias and prefix-based approaches.

cs.NE↗

arXiv2025

Shaping Inductive Bias in Diffusion Models through Frequency-Based Noise Control

Diffusion Probabilistic Models (DPMs) are powerful generative models that have achieved unparalleled success in a number of generative tasks. In this work, we aim to build inductive biases into the training and sampling of diffusion models to better accommodate the target distribution of the data to model. For topologically structured data, we devise a frequency-based noising operator to purposefully manipulate, and set, these inductive biases. We first show that appropriate manipulations of the noising forward process can lead DPMs to focus on particular aspects of the distribution to learn. We show that different datasets necessitate different inductive biases, and that appropriate frequency-based noise control induces increased generative performance compared to standard diffusion. Finally, we demonstrate the possibility of ignoring information at particular frequencies while learning. We show this in an image corruption and recovery task, where we train a DPM to recover the original target distribution after severe noise corruption.

cs.LG↗

arXiv2024

General Causal Imputation via Synthetic Interventions

Given two sets of elements (such as cell types and drug compounds), researchers typically only have access to a limited subset of their interactions. The task of causal imputation involves using this subset to predict unobserved interactions. Squires et al. (2022) have proposed two estimators for this task based on the synthetic interventions (SI) estimator: SI-A (for actions) and SI-C (for contexts). We extend their work and introduce a novel causal imputation estimator, generalized synthetic interventions (GSI). We prove the identifiability of this estimator for data generated from a more complex latent factor model. On synthetic and real data we show empirically that it recovers or outperforms their estimators.

cs.LG↗

arXiv2023

Contrastive Retrospection: honing in on critical steps for rapid learning and generalization in RL

In real life, success is often contingent upon multiple critical steps that are distant in time from each other and from the final reward. These critical steps are challenging to identify with traditional reinforcement learning (RL) methods that rely on the Bellman equation for credit assignment. Here, we present a new RL algorithm that uses offline contrastive learning to hone in on these critical steps. This algorithm, which we call Contrastive Retrospection (ConSpec), can be added to any existing RL algorithm. ConSpec learns a set of prototypes for the critical steps in a task by a novel contrastive loss and delivers an intrinsic reward when the current state matches one of the prototypes. The prototypes in ConSpec provide two key benefits for credit assignment: (i) They enable rapid identification of all the critical steps. (ii) They do so in a readily interpretable manner, enabling out-of-distribution generalization when sensory features are altered. Distinct from other contemporary RL approaches to credit assignment, ConSpec takes advantage of the fact that it is easier to retrospectively identify the small set of steps that success is contingent upon (and ignoring other states) than it is to prospectively predict reward at every taken step. ConSpec greatly improves learning in a diverse set of RL tasks. The code is available at the link: https://github.com/sunchipsster1/ConSpec

cs.LG↗

arXiv2023

Forecaster: Towards Temporally Abstract Tree-Search Planning from Pixels

The ability to plan at many different levels of abstraction enables agents to envision the long-term repercussions of their decisions and thus enables sample-efficient learning. This becomes particularly beneficial in complex environments from high-dimensional state space such as pixels, where the goal is distant and the reward sparse. We introduce Forecaster, a deep hierarchical reinforcement learning approach which plans over high-level goals leveraging a temporally abstract world model. Forecaster learns an abstract model of its environment by modelling the transitions dynamics at an abstract level and training a world model on such transition. It then uses this world model to choose optimal high-level goals through a tree-search planning procedure. It additionally trains a low-level policy that learns to reach those goals. Our method not only captures building world models with longer horizons, but also, planning with such models in downstream tasks. We empirically demonstrate Forecaster's potential in both single-task learning and generalization to new tasks in the AntMaze domain.

cs.AI↗

arXiv2023

Network Analysis of the iNaturalist Citizen Science Community

In recent years, citizen science has become a larger and larger part of the scientific community. Its ability to crowd source data and expertise from thousands of citizen scientists makes it invaluable. Despite the field's growing popularity, the interactions and structure of citizen science projects are still poorly understood and under analyzed. We use the iNaturalist citizen science platform as a case study to analyze the structure of citizen science projects. We frame the data from iNaturalist as a bipartite network and use visualizations as well as established network science techniques to gain insights into the structure and interactions between users in citizen science projects. Finally, we propose a novel unique benchmark for network science research by using the iNaturalist data to create a network which has an unusual structure relative to other common benchmark networks. We demonstrate using a link prediction task that this network can be used to gain novel insights into a variety of network science methods.

cs.SI↗

arXiv2023

A Comparison of Classical and Deep Reinforcement Learning Methods for HVAC Control

Reinforcement learning (RL) is a promising approach for optimizing HVAC control. RL offers a framework for improving system performance, reducing energy consumption, and enhancing cost efficiency. We benchmark two popular classical and deep RL methods (Q-Learning and Deep-Q-Networks) across multiple HVAC environments and explore the practical consideration of model hyper-parameter selection and reward tuning. The findings provide insight for configuring RL agents in HVAC systems, promoting energy-efficient and cost-effective operation.

cs.LG↗

arXiv2022

Towards Safe Mechanical Ventilation Treatment Using Deep Offline Reinforcement Learning

Mechanical ventilation is a key form of life support for patients with pulmonary impairment. Healthcare workers are required to continuously adjust ventilator settings for each patient, a challenging and time consuming task. Hence, it would be beneficial to develop an automated decision support tool to optimize ventilation treatment. We present DeepVent, a Conservative Q-Learning (CQL) based offline Deep Reinforcement Learning (DRL) agent that learns to predict the optimal ventilator parameters for a patient to promote 90 day survival. We design a clinically relevant intermediate reward that encourages continuous improvement of the patient vitals as well as addresses the challenge of sparse reward in RL. We find that DeepVent recommends ventilation parameters within safe ranges, as outlined in recent clinical trials. The CQL algorithm offers additional safety by mitigating the overestimation of the value estimates of out-of-distribution states/actions. We evaluate our agent using Fitted Q Evaluation (FQE) and demonstrate that it outperforms physicians from the MIMIC-III dataset.

cs.LG↗