• Write Abstention Into the Reward. Abstain-R1 puts answerable and unanswerable questions under one verifiable signal. A 3B model matches DeepSeek-R1 on three refusal benchmarks without regressing on answerable ones.

• LoRA Merge Interference Lives in the B Matrix. Pico does data-free calibration as a plug-in over Task Arithmetic, TIES, and TSV-M, adding 3.4–8.3 points on average across eight benchmarks.

• "Respect Pluralism" Becomes a Jailbreak Channel. Wrap harmful requests in moral-grayzone framing and jailbreak success rates climb across mainstream LLMs and guardrails. First time alignment objectives colliding gets used as an attack vector.

• Visual Token Compression Changes Strategy. EvoComp runs evolutionary search for per-image soft labels offline, then distills a lightweight compressor. 3x compression keeps 99.3% accuracy.

Also Notable

• HeLa-Mem Replaces Vector Retrieval With Hebbian Association — agent long-term memory becomes a connected graph, not a pile of independent vectors.
• OASIS Adds Hierarchical Event Memory to Streaming Video Reasoning — sparse evidence, unbounded redundancy, on-demand extraction instead of bigger context.
• PRISM Splits Hallucination Into Three Categories — "reasoning error / instruction drift / source-memory error" guides fixes better than output-level scoring.
• SIF Fingerprints LVLMs With In-Distribution, Uniquely Responding Samples — no OOD queries, no disruption to normal use.
• CogGen Turns Deep Research Reports Into a Recursive Nonlinear Pipeline — moves beyond the "retrieve, outline, fill" straight line.
• Local-First QA Agent for Continuous Glucose Monitoring — patient data stays on-device, aimed at CGM daily self-management.
• SAVE Adds Gene-Block Attention to Single-Cell Generation — genes aren't independent tokens anymore, multi-condition generation supported.
• Diffusion Models Get Noise-Adaptive Sampling for Inverse Problems — no task-level tuning, one sampling strategy across inverse problems.
• Schema-Level Diagnosis for Subjective NLP Annotation Disagreement — is the standard ambiguous or is diversity legitimate? A practical way to tell them apart.

Read the full edition →