Hi, I'm Arav Dhoot.

An extension tracing how hint-following and concealment emerge across OLMo-3's training stages — pretraining, SFT, DPO, and RLVR. The behaviors follow a non-monotonic trajectory, with SFT checkpoints most susceptible to incorrect hints.

A replication across 11 models from 6 families: models readily follow simple hints without verbalizing them, but revert to baseline on complex hints that require real computation — so unfaithful chains of thought stay monitorable where it matters.

Consistency training encourages models to behave similarly across different contexts, and has shown promise for reducing misalignment. We broaden the scope of consistency training in two ways. First, we introduce two new internal consistency targets: MLP Consistency Training (MLPCT), which matches post-activation MLP states, and Attention Consistency Training (AttCT), which matches per-head attention distributions. Second, we apply consistency training to four additional safety threats: persona in-context learning attacks, adversarial frustration, prefill attacks, and conditional misalignment. Across several models and threat settings, we find that consistency training reduces misalignment well beyond the sycophancy and jailbreak settings studied in prior work. We also find cases of cross-threat generalization, where training against one failure mode improves robustness to another, and identify a shared residual-stream mechanism underlying ACT, MLPCT, and AttCT, while distinguishing BCT as mechanistically distinct. Our results suggest that consistency training is a flexible and extensible framework for alignment, capable of unifying defenses against a broader class of model pathologies.

Attempts to reproduce Gemma 3's frustration behaviors on Gemma 4 using repeated rejection, prefill attacks, and reasoning-trace analysis — and fails: the new model stays markedly more stable across every vector tried.

Code transformation is a foundational capability in the software development process, where its effectiveness relies on constructing a high-quality code representation to characterize the input code semantics and guide the transformation. Existing approaches treat code transformation as an end-to-end learning task, leaving the construction of the representation needed for semantic reasoning implicit in model weights or relying on rigid compiler-level abstractions. We present SemRep, a framework that improves code transformation through generative code representation learning. Our key insight is to employ the semantics-preserving transformations as the intermediate representation, which serves as both a generative mid-training task and the guidance for subsequent instruction-specific code transformations. Across general code editing and optimization tasks (e.g., GPU kernel optimization), SemRep outperforms the extensively finetuned baselines with strictly the same training budget by 6.9% in correctness, 1.1x in performance, 13.9% in generalization, and 6.7% in robustness. With the improved exploration of diverse code transformations, SemRep is particularly amenable to evolutionary search. Combined with an evolutionary coding agent, SemRep finds optimizations that 685B larger-weight baselines fail to discover while achieving the same performance with 25% less inference compute.

A satellite architecture that puts flood detection directly in orbit: spiking neural networks fuse optical and synthetic-aperture-radar imagery onboard power-efficient neuromorphic processors (Intel Loihi 2, BrainChip Akida, 2–5W), cutting flood intelligence delivery from the usual hours down to minutes for emergency and humanitarian response.

Parkinson's Disease is a prevalent neurodegenerative disorder affecting millions worldwide, with early diagnosis remaining a critical challenge, particularly in underserved areas. This paper proposes a novel framework for the diagnosis and severity classification of PD using video data captured in uncontrolled environments. Leveraging deep learning techniques, the approach synthesizes Skeleton Energy Images (SEIs) from gait sequences and employs a CNN, a ResNet, and a Vision Transformer to analyze these images, allowing accurate detection of PD and differentiation of its severity without specialized equipment or professional oversight. The ResNet-18 model achieved 100% test accuracy, with the ViT and CNN five-layer models at 99.96% and 96.40% respectively. Given the nature of the dataset used, the framework stands to function as a cost-effective and accessible tool for early PD detection in various healthcare settings.