Dylan Jayabahu — AI researcher, CS at Waterloo.

Below-chance truth-probe scores need not mean a model has hidden truth information. In the compliant-context setup studied here, truth and action training labels coincide: fitting either probe solves the same optimization. On adversarial evaluation data, scoring that direction against opposite labels forces complementary AUROCs. Mixed-context fitting separates the two targets.

The aliasing rate then follows in closed form from two published behavioural rates and no activations at all, released as a 40-line audit function. The constructive half is a randomized-codebook task plus mixed-context fitting: on identical activations of a Gemma-2-9B policy lying on 100% of adversarial trials, the conventional probe reads 0.006 while the identified probe reads 1.000 with zero seed variance, holding at the final layer in 19 of 21 cells across four architectures (3B to 32B). Scoped to linear decodability, not causal use.

Can a reasoning model learn to stop earlier through activation-target training? This study evaluates full-vector and on-axis reconstruction across three model scales, measuring shortening alongside accuracy and termination. A 376-job development comparison selected candidate pairs at three shortening targets; held-out validation of those pairs is pending.

Historical experiments demonstrate learned shortening and cross-domain transfer, with accuracy costs in some settings. Independently selecting each objective family exposes tradeoffs that fixed-recipe comparisons miss. Off-axis drift is associated with some scalar outcomes, but its causal role and the necessity of full-vector reconstruction remain unestablished; both objectives also install refusal.

A reasoning model reaches its answer roughly twice as early as it stops writing. Found a causal “halt vector” in the residual stream and internalized it into the weights: a hook-free, per-problem self-halt that cuts overthinking without losing accuracy.

Difference-of-means halt vector at layer 18 of DeepSeek-R1-Distill-Qwen-7B. Steering it controls generation length monotonically (ρ = −0.82) where a replicated value axis is inert (+0.09). Internalized by reconstructing the full steered activation with the off-axis dimensions pinned; the scalar-projection target backfires, running 43% longer at 0.57 accuracy. Fit from 24 problems, no RL; stable across 3 seeds; closes an AIME cap-loop pathology (31–38% → 6–9%). That scalar-versus-full-vector question is taken up at scale in Off-Axis Drift, above.

Dense LoRA checkpointing turns a single GRPO run into a frame-by-frame movie of a 7B reasoning model's internals, so a failure mode can be studied as it forms, not just post-hoc.

GRPO/RLVR + dense LoRA checkpointing + per-step probing. Under a length penalty, reasoning length collapses and held-out accuracy falls — a known effect, reproduced across three historical seeds. A literature review and two self-audits retired the original headlines; what survives is descriptive: in two of three seeds a first-answer-token concentration readout ended at 0.81–0.85 while accuracy fell 43–45%, so high readout concentration did not assure correct answers. The current paper asks how GRPO’s std and length normalizations jointly change the penalty’s accuracy cost; the std effect alone is already published. That four-arm factorial is implemented and pre-registered but not yet run; the pipeline has been calibrated on cloud H200s with a 50-step baseline run. Its fixed 500-question cohort replaces the historical seed-dependent splits; the two sets of evaluations must not be pooled.

A computer-aided diagnosis pipeline for invasive ductal carcinoma that pairs sliding-window heatmaps with a homogeneous-region oversampling technique, preserving biological structure while correcting class imbalance. Evaluated on a public histopathology benchmark.

Sliding-window heatmaps visualize the model’s patch-level predictions. Evaluated on the public dataset and indexed on IEEE Xplore.

The first FMCW mmWave dataset captured in natural living environments rather than a lab setup, positioning privacy-preserving radar as a viable backbone for aging-in-place monitoring. CNN baselines establish where the real difficulty lies.

TI IWR6843AOP FMCW sensor · 3D point clouds · 3-layer CNN with sliding-window detection. Presented at CVIS 2024.