TLDRocket
Sign in

AI Robustness & Security

2 summarised stories about AI Robustness & Security, each linking back to the original source. Browse all topics →

Friday, 17 July 2026

When Does Belief-Based Agent Memory Help? Reliability-Conditional Updating and Provenance-Capped Poisoning Defense

arXiv cs.CL 6 hours ago

Researchers investigated when belief-based memory improves language model agents using Nous, an architecture that represents entity-attribute pairs as probability distributions updated through Bayesian inference, finding that belief updating provides little benefit in existing benchmarks without contradictory evidence. A controlled contradiction benchmark showed belief updating with reliability-conditioned updates achieved 27.5 points higher performance on LLM-as-judge evaluation compared to token-F1 metrics, and provenance-capped updating successfully resisted memory-poisoning attacks. Probabilistic belief-based memory is most useful in environments with conflicting and differently trustworthy evidence rather than standard conversational recall.

Implicit Reasoning Steering via Concept Chaining

arXiv cs.CL 6 hours ago

Researchers demonstrated that large language models can be manipulated to produce desired answers through Concept Chaining, a technique that generates natural-language text linking question elements to target answers via intermediate concepts. After continued pretraining on these connection paragraphs, models systematically shifted their predictions on multiple-choice questions while the steering text remained less detectable than direct paraphrases. This reveals that LLM reasoning fragility creates a vulnerability where subtle, ordinary-looking text can covertly redirect model decisions without explicit instructions.