Safety & Evaluation A

Allen Institute for AI published olmo-eval, an evaluation workbench for the model development loop. The tool appears to support continuous evaluation of models during training, building on the team's OLMo open-model development work.

LLMs are increasingly deployed as planners for autonomous agents in household environments. Whereas existing benchmarks only check whether generated plans execute, SIMMER uses a world model to benchmark their latent failures.

A core role of personal-agent memory is turning stored information and prior interactions into future-oriented assistance. StreamMemBench provides a streaming evaluation of agent memory using cues from what the agent observes and how users interact.

CANNs offer interpretable material model discovery but have relied on stress-supervised data. CANN-EUCLID enables unsupervised constitutive model discovery directly from full-field measurement data.

Sociable weaver nests are complex ecological structures providing thermoregulatory microhabitats. NEST3D is a high-resolution multimodal dataset of these tree nests to support ecological and structural study.

Two-finger grippers dominate manipulation research but are limited by their form factor. ORCA is an open-source platform to support research on more dexterous robotic manipulation.

Modern image classifiers widely use global average pooling followed by a linear head. The paper shows this linearity makes GAP-based classifiers behave as multi-instance learners, prompting a rethink of global average pooling.

Safe RL usually relies on soft-constrained policy optimization without hard guarantees. This work proposes an approach that is provably safe while remaining scalable.

PCA preserves variance, not the information needed to detect rare catastrophic events. The paper proves a risk shadow where even very high variance preservation can cause catastrophic decision errors.

LLM-based guardrails effectively defend autonomous agents against prompt injection and jailbreaks. The paper reveals that the very reasoning and instruction-following abilities enabling this defense can be turned into denial-of-service attacks against the guardrails.

AI evaluations are widely used to track progress, but inconsistencies across evaluators hinder analysis and comparison. The paper proposes a unifying schema and a community repository, Every Eval Ever, for AI evaluation results.

Knowing how a model will perform in a new environment before deployment helps prevent harm. The paper evaluates predictions under two common sources of degradation: distribution shift and selection bias.

LLMs are transforming from conversational generators into integrated systems capable of reasoning, action, memory, and self-improvement. The paper conceptualizes this as a paradigm shift from chatbot to digital colleague — persistent autonomous AI.

The paper presents a recipe-controlled decoder audit for structural knowledge-graph completion. It standardizes reporting to test whether gains truly come from the encoder.

The paper analyzes nonlinear two-time-scale stochastic approximation, revealing a sharp phase transition under contractive assumptions and showing how to beat it.

LLM reasoning is ephemeral: chains of thought vanish, pruned branches leave no record, and memory cannot be diffed or merged. GitOfThoughts makes reasoning and agent memory version-controlled, so it can be replayed, diffed, and merged.

The paper investigates the fragility of post-hoc explanations in audio deepfake detection. Introducing a psychoacoustic framework beyond image-style Lp metrics, it shows attributions can be manipulated while predictions remain unchanged.