• Context Shifts Fail to Weaken Occupational Stereotypes. ContextBias tests 92 occupations across 66,240 images. Irrelevant settings increase attribute concentration across occupations by 0.047.

• Functional Vectors Transfer Emotion Recognition Across Languages. Zero-shot injection partially reproduces few-shot prompting, though language distance and cultural differences still need testing.

• ACTD Moves Reasoning Across Tokenizers. Anchor alignment and residual regularization offer a path for compressing heterogeneous models. The gains and costs of multiple teachers require closer review.

• Hardware Faults May Damage Video Semantics Before Image Quality. A single fault cuts overall performance by up to 3.7%. Memory and high-order exponent bits deserve extra protection.

Also Notable

• Separate Safety Refusals From Social Bias. Strong guardrails can trigger refusals that obscure a vision-language model's underlying bias. A new method measures the two behaviors separately..
• Compare Text Anonymization Across Explicit Goals. PrivBench provides a modular platform for evaluating text-to-text privacy methods without relying on one metric..
• Static Semantic IDs Lose Query Intent. ICEGR tries to preserve intent throughout the training pipeline for generative e-commerce retrieval..
• Choose QAT Methods by Deployment Goal. This survey maps error profiles, numerical formats, strategy transfer, and deployment evaluation across quantization objectives..
• Resolve References Beyond Conversation History. Understanding “the fix we discussed yesterday” requires locating indirect references in workspaces and tool outputs. A new benchmark tests this capability in collaborative agents..
• Combine Skills With Formal Verification. SkillForge uses verifier feedback to refine composable skills for constrained Dafny program generation..
• Preserve Alignment Through the Generation Path. Random masking in discrete diffusion breaks spatial correspondence. DDB uses diffusion bridges to retain alignment for image translation and generation..
• Place Audio Events on the Timeline. TEMPO uses multitask post-training to improve temporal grounding for events, speakers, and sounds..
• Study World Models and Policies Together. Spectral and behavioral analyses reveal how environment understanding combines with task execution..
• Make Spatial Audio Obey Physics. PhysWave adds physical guidance to latent diffusion so direction and acoustic relationships remain testable..

Read the full edition →