COLM 2026

Expanded account

Registration and presentation history — 2026.

This record covers Ethosoft's 2026 conference season, registered under COLM 2026 (the Conference on Language Modeling). Three workshop-event presentations from this season are listed below, drawn from the team's CV records; each targets a different research line while sharing an evaluation-first approach to claims.

Presentation history

• UGTC: Uncertainty-Gated Temporal Credit — Adaptive Advantage Estimation for Continually Learning Agents. A modular advantage estimator for actor-critic reinforcement learning that blends fast and slow critics through an uncertainty-based gating mechanism, allowing the bias–variance tradeoff to be tuned per state; it integrates directly with PPO, TD3, SAC, and DreamerV3 through architecture-specific hooks. Accepted (double-blind review) at the ExWBC workshop, RSS 2026 — the 22nd Robotics: Science and Systems conference, Sydney, Australia.
• Removal Cost as a Measurement Axis for Open-Weight Language Model Safety. Proposes evaluating safety mechanisms in open-weight language models by the cost of removing them, rather than only confirming their presence—testing whether a safety component is genuinely load-bearing by measuring how much the system degrades once it is taken out. Part of the MYCELIUM research line. Accepted (double-blind review, at least three independent reviewers) at AIMS: AI Measurement Science Workshop, co-located with COLM 2026, Hilton Union Square, San Francisco, USA.
• Efficient Rewarding of Agent Plans Without LLM-as-Judge: Goal Drift Score for GRPO. Related to Ethosoft's Goal Drift research on how AI systems revise, preserve, or reinterpret objectives when task specifications are incomplete or exposed to changing context; read the full paper record for details.