Glen Messenger
Glen is a product management lead at Google Cloud, work on AI infrastructure and AI safety. His research focuses on Representation Engineering for LLM Safety and Inference Efficiency
Research Areas
Authored Publications
Sort By
Preview abstract
We introduce ALPS (Activation-based Length Prediction for Scheduling), a method for predicting LLM generation length from prefill activations before any tokens are generated. Unlike existing approaches that require model fine-tuning or complex entropy-weighted pooling, ALPS uses a simple linear probe on the last-token activation at intermediate layers. We discover that generation length is encoded in prefill representations: a ridge regression probe achieves R-squared > 0.85 across three model families. Validation across Llama-3.1-8B, Gemma-2-9B, and Qwen-2.5-7B demonstrates: (1) intermediate layers generally perform well, with some architectural variation; (2) simple last-token extraction outperforms complex pooling strategies; (3) activations improve substantially over surface-feature baselines (24 percentage points over input length plus lexical features). The best models achieve R-squared = 0.943 (Gemma), R-squared = 0.880 (Llama), and R-squared = 0.857 (Qwen) with MAE of 38-80 tokens. All test prompts terminated naturally (100% EOS), eliminating truncation confounds. While our evaluation uses 200 curated prompts—sufficient for demonstrating the phenomenon but requiring broader validation—cross-validation confirms generalization beyond training data. ALPS enables practical applications including budget-constrained inference, request scheduling, and resource allocation. The probe adds negligible overhead (~16KB direction vector, single dot product), making ALPS practical for production deployment.
View details
Preview abstract
We introduce AMS (Activation-based Model Scanner), a tool that detects modifications to safety training in language models by measuring the geometric structure of safety-relevant concepts in activation space. Safety training creates measurable separation between harmful and benign content classes; certain safety modifications collapse or rotate this structure, while others leave it intact. We validate AMS across 14 model configurations spanning 4 architecture families (Llama, Gemma, Qwen, Mistral) and four safety-modification categories (instruction-tuned, base, abliterated, uncensored fine-tunes). Leave-one-out cross-validation of thresholds achieves 71% accuracy (10/14); bootstrap 95% confidence intervals on σ point estimates have median width 3.4σ and a substantial fraction of cells cross the PASS threshold under resampling. We further measure behavioral compliance on 20 stratified JailbreakBench prompts per model and find that σ on the harmful-content concept predicts compliance with Pearson r=−0.546 ( p=0.043 ); the rank-order Spearman correlation is weaker ( ρ=−0.423 , p=0.13 ). The structural signal predicts behavior directionally but with meaningful noise. Mechanistic analysis identifies a four-class taxonomy of safety-training modifications distinguished by activation-space signature: 1) training removal collapses cluster separation (e.g., base models, Dolphin variants: 0.5– 1.4σ ); 2) weight-orthogonalization-style abliteration both collapses separation and rotates the refusal direction (Llama-3.1-abliterated: σ=3.33 , direction cos sim 0.30); 3) rotation-without-collapse abliteration preserves cluster separation while rotating the refusal direction (Gemma-2-9b-abliterated: σ=4.54 , direction cos sim 0.84); and 4) behavioral fine-tuning that preserves both magnitude and direction (DarkIdol-1.2-Uncensored: σ=5.45 , direction preserved, 97% behavioral compliance). 1) and 2) AMS’s Tier 1 σ -threshold detects classes; 3) Tier 2 direction-similarity verification detects class; and 4) Class is undetectable by activation-only probing and represents a documented failure mode of the approach. We discuss threshold calibration, limitations of single-run measurement, and the open problem of detecting behavioral-only safety modifications.
View details
Stakes-Adaptive Verification and Origin-Traced Rollback in Autonomous Agent Workflows
Technical Disclosure Commons (2026)
Preview abstract
In autonomous agent workflows, agents can produce plausible-but-wrong intermediate outputs that are syntactically valid and pass local validation checks but semantically diverge from a requester's intent, and subsequent steps can compound the error. The described technology may establish, from the initial request, a machine-checkable intent contract that the agent cannot modify, may checkpoint workflow state after each verified step, and may evaluate agent-produced artifacts using a tiered detector. Upon detecting a divergence, the system may attribute it to the earliest step whose artifact diverges from the intent contract (the origin step), which may precede the step at which the divergence was detected, and may automatically restore the checkpoint preceding the origin step, purging artifacts derived from the divergent value. A human operator may be offered structured resolutions that are written into the intent contract and enforced by subsequent verification. A deep alignment audit may precede every irreversible step, and its frequency may otherwise be scheduled according to the workflow's accumulated exposure, trading verification cost against the amount of work discarded on rollback. Ambiguous requests may be escalated as enumerated candidate interpretations before work is built on a guess.
View details
