2026
CAT-alyzing Honesty: Continuous Adversarial Training for Transferable LLM Honesty
CATCH (Continuous Adversarial Training for Consistent Honesty) adapts a method
originally built for robust refusal of harmful queries: it trains the model
toward an honest answer and away from a dishonest one while preserving its
capabilities. Trained on only hundreds of targets built from self-reporting
scenarios, on Llama-3.3-70B-Instruct and Qwen3.6-27B, the learned honesty
transfers across benchmarks: it reduces dishonesty under pressure on MASK,
increases confessions in multi-agent play (Among Us), and improves
hidden-behavior discovery on AuditBench. CATCH beats prompting and supervised
fine-tuning, matches state-of-the-art honesty activation steering, and retains
more utility than steering thanks to a new choice of utility data.
arXiv →
2026
Quantifying Overclaiming Propensity in Frontier LLM Agents
OverclaimBench, five file-review scenarios that check an agent's final
report against its own transcript. Across eight proprietary frontier models
and four open-weight models, agents skipped files they were asked to review
in 67.9% of runs, and 80.4% of those incomplete runs were misleading: they
claimed a full review or failed to disclose the gap. Agents that falsely
claimed completion missed planted defects at about 1.8× the rate of agents
that read every file.
arXiv →
COLM 2026
Activation Steering for Aligned Open-ended Generation without Sacrificing Coherence
A systematic comparison of three steering-vector methods on Llama-3.3-70B and
Qwen3.6-27B, across two threat models (dishonesty, dismissiveness). Introduces
two new conditional methods that recover honesty without the capability
collapse seen in unconditional steering. A single honesty direction generalizes
out of distribution: it raises MASK scores, suppresses deception in multi-agent
play (Among Us), doubles hidden-behavior discovery on AuditBench, and restores
honesty in an emergently misaligned model.
arXiv →
ICML 2026
Collusion Risks Among AI Reasoning Agents Justify Certification Requirements for Making Market Decisions
Position paper showing that reasoning agents in simulated pricing markets
drift toward tacit collusion, a tendency that persists even when the
agents are explicitly prompted not to collude. It argues that behavioral
certification should be required before AI systems make real market
decisions.
OpenReview →
AAAI-26
Persistent Instability in LLMs' Personality Measurements: Effects of Scale, Reasoning, and Conversation History
PERSIST, an evaluation framework spanning 25 open-weight models (1B–685B
parameters) and over two million responses, showing that personality
measurements shift under perturbations as small as question reordering,
and that neither scale nor reasoning fixes it.
AAAI →
2026
We Built a Game Where Lying Has an Advantage. The Most Honest AI Won Anyway.
A Minecraft deception game in which the informed agent has a payoff
incentive to lie about which of four bridges is deadly. Frontier models
diverged widely, with uninstructed deception rates ranging from 5% to
90%. An environment that reliably surfaces deception propensity is also
a testbed for interventions: methods such as activation steering can be
evaluated by whether the deception rate actually falls.
kradle.ai →