A transformer architecture alteration to incentivise externalised reasoning

We propose a new architectural change, and post-training pipeline, for making LLMs more verbose reasoners by teaching a model to truncate forward passes early. We augment an existing transformer architecture with an early-exit mechanism at intermediate layers and train the model to exit at shallower layers when the next token can be predicted without deep computation. After a calibration stage, we incentivise the model to exit as early as possible while maintaining task performance using reinforcement learning. We provide preliminary results to this effect for small reasoning models, showing that they learn to adaptively reduce computations across tokens. We predict that, applied at the right scale, our approach can minimise the amount of excess computation that reasoning models have at their disposal to perform non-myopic planning using their internal activations, reserving this only for difficult-to-predict tokens.

Diagnosing Pathological Chain-of-Thought in Reasoning Models

Chain-of-thought (CoT) reasoning is fundamental to modern LLM architectures and represents a critical intervention point for AI safety. However, CoT reasoning may exhibit failure modes that we note as pathologies, which prevent it from being useful for monitoring. Prior work has identified three distinct pathologies: post-hoc rationalization, where models generate plausible explanations backwards from predetermined answers; encoded reasoning, where intermediate steps conceal information within seemingly interpretable text; and internalized reasoning, where models replace explicit reasoning with meaningless filler tokens while computing internally. To better understand and discriminate between these pathologies, we create a set of concrete metrics that are simple to implement, computationally inexpensive, and task-agnostic. To validate our approach, we develop model organisms deliberately trained to exhibit specific CoT pathologies. Our work provides a practical toolkit for assessing CoT pathologies, with direct implications for training-time monitoring.

Chain-of-thought obfuscation learned from output supervision can generalise to unseen tasks

Chain-of-thought (CoT) reasoning provides a significant performance uplift to LLMs by enabling planning, exploration, and deliberation of their actions. CoT is also a powerful tool for monitoring the behaviours of these agents: when faithful, they offer interpretations of the model's decision making process, and an early warning sign for dangerous behaviours. However, optimisation pressures placed on the CoT may cause the model to obfuscate reasoning traces, losing this beneficial property. We show that obfuscation can generalise across tasks; models that learn to obfuscate reasoning involving reward hacking (e.g. accessing and utilising leaked information) generalise both the reward hacking behaviour and its obfuscation in CoT to unseen reward hacking settings. Most worryingly, we show that obfuscation of CoT reasoning, and its generalisation across tasks, also follows when we penalise only the model's final actions after closing its CoT. Our findings suggest that current practices of penalising harmful generations may inadvertently lead to a reduction in the broader monitorability of LLMs in unpredictable ways.

Alignment Pretraining: AI Discourse Causes Self-Fulfilling (Mis)alignment

Pretraining corpora contain extensive discourse about AI systems, yet the causal influence of this discourse on downstream alignment remains poorly understood. If prevailing descriptions of AI behaviour are predominantly negative, LLMs may internalise corresponding behavioural priors, giving rise to self-fulfilling misalignment. This paper provides the first controlled study of this hypothesis by pretraining 6.9B-parameter LLMs with varying amounts of (mis)alignment discourse. We find that discussion of AI contributes to misalignment. Upsampling synthetic training documents about AI misalignment leads to a notable increase in misaligned behaviour. Conversely, upsampling documents about aligned behaviour reduces misalignment scores from 45% to 9%. We consider this evidence of self-fulfilling alignment. These effects are dampened, but persist through post-training. Our findings establish the study of how pretraining data shapes alignment priors, or alignment pretraining, as a complement to post-training. We recommend practitioners pretrain for alignment as well as capabilities. Our models and datasets are available at alignmentpretraining.ai

A flexible Bayesian non-parametric mixture model reveals multiple dependencies of swap errors in visual working memory

Human behavioural data in psychophysics has been used to elucidate the underlying mechanisms of many cognitive processes, such as attention, sensorimotor integration, and perceptual decision making. Visual working memory has particularly benefited from this approach: analyses of VWM errors have proven crucial for understanding VWM capacity and coding schemes, in turn constraining neural models of both. One poorly understood class of VWM errors are swap errors, whereby participants recall an uncued item from memory. Swap errors could arise from erroneous memory encoding, noisy storage, or errors at retrieval time - previous research has mostly implicated the latter two. However, these studies made strong a priori assumptions on the detailed mechanisms and/or parametric form of errors contributed by these sources. Here, we pursue a data-driven approach instead, introducing a Bayesian non-parametric mixture model of swap errors (BNS) which provides a flexible descriptive model of swapping behaviour, such that swaps are allowed to depend on both the probed and reported features of every stimulus item. We fit BNS to the trial-by-trial behaviour of human participants and show that it recapitulates the strong dependence of swaps on cue similarity in multiple datasets. Critically, BNS reveals that this dependence coexists with a non-monotonic modulation in the report feature dimension for a random dot motion direction-cued, location-reported dataset. The form of the modulation inferred by BNS opens new questions about the importance of memory encoding in causing swap errors in VWM, a distinct source to the previously suggested binding and cueing errors. Our analyses, combining qualitative comparisons of the highly interpretable BNS parameter structure with rigorous quantitative model comparison and recovery methods, show that previous interpretations of swap errors may have been incomplete.

Who Needs Decoders? Efficient Estimation of Sequence-level Attributes

State-of-the-art sequence-to-sequence models often require autoregressive decoding, which can be highly expensive. However, for some downstream tasks such as out-of-distribution (OOD) detection and resource allocation, the actual decoding output is not needed just a scalar attribute of this sequence. In these scenarios, where for example knowing the quality of a system's output to predict poor performance prevails over knowing the output itself, is it possible to bypass the autoregressive decoding? We propose Non-Autoregressive Proxy (NAP) models that can efficiently predict general scalar-valued sequence-level attributes. Importantly, NAPs predict these metrics directly from the encodings, avoiding the expensive autoregressive decoding stage. We consider two sequence-to-sequence task: Machine Translation (MT); and Automatic Speech Recognition (ASR). In OOD for MT, NAPs outperform a deep ensemble while being significantly faster. NAPs are also shown to be able to predict performance metrics such as BERTScore (MT) or word error rate (ASR). For downstream tasks, such as data filtering and resource optimization, NAPs generate performance predictions that outperform predictive uncertainty while being highly inference efficient.