Dylan Jayabahu — AI researcher, CS at Waterloo.

Truth probes fall below chance on models trained to deceive, and the field reads that as the model hiding what it knows. It cannot be. On the compliant data those probes are fit on, belief, intent, and emitted token are one label vector, so a “truth probe” and an “action probe” are a single fitted direction scored against opposite labels. Nothing had to be hidden.

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.

Interpretability keeps producing read-then-steer directions (refusal axes, value axes, truthfulness features) that only work at inference, behind a hook. Moving one into the weights turns out to demand the whole steered activation: matching its scalar projection, the target prior work adopted untested, does not install the behaviour at any scale tested.

Across Qwen 1.5B/7B/14B, no matched-capacity scalar arm compresses at held accuracy, while full-vector reconstruction cuts 22.6–33.1%. The mechanism is off-axis drift: satisfying one scalar leaves every other dimension free, and the frozen downstream reader depends on those. The direction itself is given a mechanistic identity, since projecting it out adds ~1,480 think-tokens and collapses termination by 23 points, its cosine overlap with the end-of-think logit direction is 0.004 (so it is not a disguised “emit </think>” push), and the effect is ~96% reader-mediated and localized to the final two MLP layers. A pre-registered second-direction test then bounded the claim: a near-orthogonal refusal direction installed under both targets, so the universality claim was struck from the main text rather than softened.

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 caught as it crystallizes, not just post-hoc.

GRPO/RLVR + dense LoRA checkpointing + per-step probing. Reasoning-length compression leads the accuracy collapse by 15–30 steps: an early-warning signal the model’s own confidence misses (it stays ~0.9 as accuracy halves). An α-sweep maps a tunable critical penalty: a phase boundary, not a ramp.

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. State-of-the-art on the public benchmark.

Interpretable heatmap visualization makes the model a clinical-grade second reader for pathologists rather than a black box. 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.