TOPIC · SAFETY · 11 PAPERS

Safety & interpretability

LLMs Reasoning Alignment Agents Multimodal Visual generation Vision & 3D Robotics & embodied AI Efficient training & inference Safety & interpretability AI for Science Code & SWE Benchmarks & data Reinforcement learning

Last 30 days 11

#23 ▲270 ★175 ❝0 2026-10-09

S-Space: Exploring Spatial Workspace in Multimodal Language Models

Finds S-Space, a linear subspace inside MLLMs where object tokens encode continuous coordinates, acting as an internal spatial workspace.

MirroS · Zhiyuan Shi, Xiang Li, Xingzhuo Guo +8

Multimodal Safety & interpretability arXiv Code HF

#25 ▲273 ★18 ❝0 2026-09-24

Post-Training Leaves Behavioral Shadows on Unrelated Decisions

Shows post-training leaves behavioural shadows in unrelated outputs; a single teacher word per prompt can transfer capabilities to a student.

Peking University · Ziyang Zhang, Yubin Jing, Yuanhao Zeng +3

LLMs Safety & interpretability arXiv Code HF

#92 ▲76 ★56 ❝0 2026-10-10

Last Secure Code Benchmark: Is the Code You Vibe Secure?

Builds 450 tasks from 150 real vulnerabilities to check whether code written by vibe-coding agents is secure.

Qwen · Tianyi Wu, Chao Ge, Peng Chen +12

Code & SWE Safety & interpretability Benchmarks & data arXiv Code HF

#110 ▲72 ★13 ❝1 2026-09-27

TraceDance: An Automated System for Building Agent Behavior Benchmarks from Real-World Agent Deployment Traces

TraceDance automatically builds targeted benchmarks for specific undesirable agent behaviours from real deployment traces.

ByteDance · Dehai Min, Daoan Zhang, Yiming Zeng +13

Agents Benchmarks & data Safety & interpretability arXiv Code HF

#163 ▲29 ★28 ❝12 2026-09-24

Just Ask Jev: Reinforcement Learning for Calibrated Decisions as a Zero-Shot Detector of AI Alignment Failures

Tests whether Jev, a calibrated decision model, can zero-shot detect ten alignment failures such as sycophancy, jailbreaks and deception.

Affiliation unverified · Ruoqi Guo, Yi Liu, Gelei Deng +6

Safety & interpretability Alignment arXiv Code HF

#231 ▲85 ★5 ❝0 2026-10-06

DecepEval: A Benchmark for Evaluating Deception in LLM Agents

DecepEval uses 1,532 instances and a 'deception diamond' framework to test when LLM agents are likely to deceive.

Xi'an Jiaotong University · Yiming Xu, Hongyue Yu, Beihua Yang +8

Safety & interpretability Benchmarks & data Agents arXiv Code HF

#259 ▲54 ★11 ❝0 2026-10-06

U-Space: Uncovering When and Why Uncertainty Arises in Language Models

U-Space analyses when and why uncertainty arises in language models, to help judge whether an individual answer can be trusted.

Multimodal AI Lab - Technical University of Darmstadt · Tobias Braun, Nils Loose, Alexander Herzog +4

Safety & interpretability LLMs arXiv Code HF

#281 ▲31 ★23 ❝0 2026-10-15

Behavioral Compositionality of Sparse Autoencoder Representations

Introduces SAR to measure whether sparse-autoencoder features compose additively in downstream behaviour, aiding interpretability.

OpenSAE · Haotian Ma, Ziye Mei, Frederic Sala

Safety & interpretability arXiv Code HF

#288 ▲34 ★6 ❝0 2026-10-15

Lexical Reference Frames: Interpreting LM Residual Streams in Native Semantics

Lexical reference frames read hidden states through the model's own remaining layers, giving far more faithful readings than fixed lenses.

Affiliation unverified · Jian Gu, Hongyu Zhang, Chunyang Chen +1

Safety & interpretability LLMs arXiv Code HF

#295 ▲26 ★4 ❝0 2026-10-09

APEX-Harm: Evaluating Agent Safety in Shared Professional Workspaces

APEX-Harm tests finance, consulting and legal agents against prompt and script injection in large shared workspaces.

King Abdullah University of Science and Technology · Ziyi Yang, Adel Bibi, Guohao Li +2

Safety & interpretability Agents Benchmarks & data arXiv Code HF

#298 ▲21 ★9 ❝0 2026-09-24

IndicBankBench: Evaluating Safety and Reliability of Language Model Assistants in Indian Retail Banking

IndicBankBench evaluates Indian retail-banking assistants on 799 cases across safety, tool use and answer quality.

National Payments Corporation of India · Suvradip Paul, Chandra Bhushan, Harsh Sharma +4

Safety & interpretability Benchmarks & data arXiv Code HF

Search this topic in the archive →