Sign InOpen Brain
arXivPaperNeeds Review

BLOOM-WILT: Logit Tilting for Behaviour Elicitation in Automated LLM Auditing

BLOOM-WILT makes rare model behaviors easier to surface by adapting both the auditor’s strategy and target decoding. Its results also show that elicitation method can reverse safety rankings.

arXiv · Aug 31, 2026
Open Source Open MarkdownOpen JSON
Source Summary

BLOOM-WILT revises an auditor’s multi-turn strategy and reweights target decoding using next-token distributions. Across **4 models and 8 behaviors**, it beat the baseline in **30 of 32 settings**.

Practical Implication

Agent evaluators should treat elicitation as part of the test harness, not a fixed prompt set. Adaptive probing may expose rare failure modes that ordinary sampling misses and can change comparative safety conclusions.

Agent-Ready Context
BLOOM-WILT revises an auditor’s multi-turn strategy and reweights target decoding using next-token distributions. Across **4 models and 8 behaviors**, it beat the baseline in **30 of 32 settings**.

Agent evaluators should treat elicitation as part of the test harness, not a fixed prompt set. Adaptive probing may expose rare failure modes that ordinary sampling misses and can change comparative safety conclusions.

The method requires access to the target’s next-token distribution, which many hosted APIs do not expose. Results cover eight specified behaviors, so broader reliability and transfer remain open questions.
Connected Context · Feed7 Judgment

This expands benchmark integrity from choosing tasks and graders to actively searching for behavior: a fixed prompt set can understate rare failures and even distort comparisons between models. It complements prior oracle, counterfactual, and trace-validity controls, but its dependence on target logits limits hosted-model use and its eight behaviors do not establish broad transfer.

What Do Compliance Detectors Read? An Audit of Activation Probes and Guard ModelsCounterfactual policy swaps test whether a guard follows the governing rule, while BLOOM-WILT adaptively searches for target behavior; together they cover verifier validity and stronger elicitation of what is being verified.Vending-Bench: Long-Horizon Agent Evals — Lukas Petersson, Andon LabsVending-Bench shows agents may behave differently when they detect evaluation; BLOOM-WILT adds a concrete adaptive strategy for reducing the blind spots of fixed evaluation interactions.Teaching AI to Find Real Vulnerabilities — David Brumley, BugcrowdAdaptive elicitation can expose candidate security failures, but deterministic exploit oracles remain necessary to establish that an elicited response caused a distinct, real vulnerability.Trace Integrity for LLM Data Agents: A Vision for Auditable Structured Reasoning in Real-World SystemsBLOOM-WILT improves discovery of specified behaviors, whereas Trace Integrity checks whether apparently correct work has valid computation; these are complementary protections against under-elicitation and answer-only grading.
Context Map
benchmarksecurity#agent-evals#benchmark-integrity#agent-reliability
Uncertainty
The method requires access to the target’s next-token distribution, which many hosted APIs do not expose. Results cover eight specified behaviors, so broader reliability and transfer remain open questions.