The Inference Report

Automatic parallelization remains a challenging problem in software engineering, particularly in identifying code regions where loops can be safely executed in parallel on modern multi-core architectures. Traditional static analysis techniques, such as dependence analysis and polyhedral models, often struggle with irregular or dynamically structured code. In this work, we propose a Transformer-based approach to classify the parallelization potential of source code, focusing on distinguishing independent (parallelizable) loops from undefined ones. We adopt DistilBERT to process source code sequences using subword tokenization, enabling the model to capture contextual syntactic and semantic patterns without handcrafted features. The approach is evaluated on a balanced dataset combining synthetically generated loops and manually annotated real-world code, using 10-fold cross-validation and multiple performance metrics. Results show consistently high performance, with mean accuracy above 99\% and low false positive rates, demonstrating robustness and reliability. Compared to prior token-based methods, the proposed approach simplifies preprocessing while improving generalization and maintaining computational efficiency. These findings highlight the potential of lightweight Transformer models for practical identification of parallelization opportunities at the loop level.

Chain-of-Thought (CoT) monitoring, in which automated systems monitor the CoT of an LLM, is a promising approach for effectively overseeing AI systems. However, the extent to which a model's CoT helps us oversee the model - the monitorability of the CoT - can be affected by training, for instance by the model learning to hide important features of its reasoning. We propose and empirically validate a conceptual framework for predicting when and why this occurs. We model LLM post-training as an RL environment where the reward decomposes into two terms: one term depending on final outputs and another term depending on the CoT. Our framework allows us to classify these two terms as "aligned", "orthogonal", or "in-conflict" before training. We predict that training with in-conflict terms will reduce monitorability, orthogonal terms will not affect it, and aligned terms will improve it. To validate our framework, we use it to classify a set of RL environments, train LLMs within those environments, and evaluate how training affects CoT monitorability. We find that (1) training with "in-conflict" reward terms reduces CoT monitorability and (2) optimizing in-conflict reward terms is difficult.

This study investigates the use of NeuralUCB for cost-aware large language model (LLM) routing. Existing routing approaches can be broadly grouped into supervised routing methods and partial-feedback methods, each with different tradeoffs in efficiency and adaptivity. We implement a NeuralUCB-based routing policy and evaluate it on RouterBench under a simulated online setting. Experimental results show that the proposed method consistently outperforms random and min-cost baselines in utility reward. Compared with the max-quality reference, our method achieves substantially lower inference cost while maintaining competitive reward. These findings suggest that NeuralUCB is a promising approach for cost-aware LLM routing, while also highlighting remaining challenges in action discrimination and exploration.

The pursuit of reducing the memory footprint of the self-attention mechanism in multi-headed self attention (MHA) spawned a rich portfolio of methods, e.g., group-query attention (GQA) and multi-head latent attention (MLA). The methods leverage specialized low-rank factorizations across embedding dimensions or attention heads. From the point of view of classical low-rank approximation, these methods are unconventional and raise questions of which objects they really approximate and how to interpret the low-rank behavior of the resulting representations. To answer these questions, this work proposes a generalized view on the weight objects in the self-attention layer and a factorization strategy, which allows us to construct a parameter efficient scheme, called Tucker Attention. Tucker Attention requires an order of magnitude fewer parameters for comparable validation metrics, compared to GQA and MLA, as evaluated in LLM and ViT test cases. Additionally, Tucker Attention~encompasses GQA, MLA, MHA as special cases and is fully compatible with flash-attention and rotary position embeddings (RoPE). This generalization strategy yields insights of the actual ranks achieved by MHA, GQA, and MLA, and further enables simplifications for MLA.

Human talkers often address listeners with language-comprehension challenges, such as hard-of-hearing or non-native adults, by globally slowing down their speech. However, it remains unclear whether this strategy actually makes speech more intelligible. Here, we take advantage of recent advancements in machine-generated speech allowing more precise control of speech rate in order to systematically examine how targeted speech-rate adjustments may improve comprehension. We first use reverse-correlation experiments to show that the temporal influence of speech rate prior to a target vowel contrast (ex. the tense-lax distinction) in fact manifests in a scissor-like pattern, with opposite effects in early versus late context windows; this pattern is remarkably stable both within individuals and across native L1-English listeners and L2-English listeners with French, Mandarin, and Japanese L1s. Second, we show that this speech rate structure not only facilitates L2 listeners' comprehension of the target vowel contrast, but that native listeners also rely on this pattern in challenging acoustic conditions. Finally, we build a data-driven text-to-speech algorithm that replicates this temporal structure on novel speech sequences. Across a variety of sentences and vowel contrasts, listeners remained unaware that such targeted slowing improved word comprehension. Strikingly, participants instead judged the common strategy of global slowing as clearer, even though it actually increased comprehension errors. Together, these results show that targeted adjustments to speech rate significantly aid intelligibility under challenging conditions, while often going unnoticed. More generally, this paper provides a data-driven methodology to improve the accessibility of machine-generated speech which can be extended to other aspects of speech comprehension and a wide variety of listeners and environments.

Current autonomous AI agents, driven primarily by Large Language Models (LLMs), operate in a state of cognitive weightlessness: they process information without an intrinsic sense of network topology, temporal pacing, or epistemic limits. Consequently, heuristic agentic loops (e.g., ReAct) can exhibit failure modes in interactive environments, including excessive tool use under congestion, prolonged deliberation under time decay, and brittle behavior under ambiguous evidence. In this paper, we propose the Triadic Cognitive Architecture (TCA), a unified mathematical framework that grounds machine reasoning in continuous-time physics. By synthesizing nonlinear filtering theory, Riemannian routing geometry, and optimal control, we formally define the concept of Cognitive Friction. We map the agent's deliberation process to a coupled stochastic control problem where information acquisition is path-dependent and physically constrained. Rather than relying on arbitrary heuristic stop-tokens, the TCA uses an HJB-motivated stopping boundary and instantiates a rollout-based approximation of belief-dependent value-of-information with a net-utility halting condition. Through empirical validation in a simulated Emergency Medical Diagnostic Grid (EMDG), we demonstrate that while greedy baselines over-deliberate under latency and congestion costs, the triadic policy reduces time-to-action while improving patient viability without degrading diagnostic accuracy in this environment.