AI Research Pulse

This week's standout papers include RLMF, AutoMem, and the Red Queen Gödel Machine, offering practical insights for ML engineers and product teams.

Microsoft's HARC adapters couple harmfulness and refusal directions to strengthen refusal robustness against jailbreaks while curbing over-refusal.

-...

Only three of 22 top ADMET models (CaliciBoost, MapLight, MapLight+GNN) reproduced reported performance; the rest failed due to missing code,...

Self-Distillation Zero turns binary rewards into dense supervision through self-revision, simplifying reward modeling in RL-based LLM training while...

The AI-2027 scenario's prediction of super-exponential progress without new architectures has not materialized, with metrics like time horizons...

Method Advances

• 🔥 STAR-KV KV Cache Compression: Dnotitia released STAR-KV, a low-rank KV cache compression method achieving up to 20x...

• Research depth: 315 AI × Bio papers from ICML 2026 curated across protein design, drug discovery, and genomics
• Diagnostic edge: AI now outperforms...

A principled approach decouples receptive fields and tokenization from pixel indices, letting U-Nets and ViTs trained at one grid resolution run effectively at others. This enables practical neural operators for scientific ML without retraining.

New offline-to-online RL method reduces sample costs when fine-tuning pre-trained policies for decision-making tasks like robotics, autonomous driving, and games.

STAR-KV achieves up to 20x KV cache compression and 6.9x attention speedup through low-rank compression plus quantization and custom GPU kernels. This...

Current alignment methods (RLHF, Constitutional AI, mechanistic interpretability) only inspect propositional content and therefore miss pre-rational...

Google's TabFM lets ML engineers run zero-shot classification and regression on structured tables in a single forward pass through frozen weights,...

InstanceControl achieves precise multi-instance control in complex image generation by using VLMs to automatically map text descriptions to visual...

AutoMem treats memory management as a learnable cognitive skill, letting LLMs decide file-system actions via automated structure revision and...

A new Logit-Contribution Scoring (LOCOS) method identifies attention heads that synthesize answers from meaning rather than literal copying in...

A new Task-Agnostic Pretraining (TAP) framework separates learning physical movement from task semantics in Vision-Language-Action models. By first...

ADOP-NAS delivers a practical evolutionary NAS approach that combines chaos-driven initialization and adaptive mutation with depthwise separable...

Safety & Auditing Tools

• 🔥 Guardrail Behavioral Monitoring: A new black-box method detects guardrail presence with 100% accuracy and...

A new black-box method detects guardrail presence in LLM systems with 100% accuracy by monitoring HTTP, lexical, and timing differences between benign...

DuoMem's dual-space distillation lifts a 4B model from 4.3% to 77.9% success on ALFWorld—nearly closing the gap to its 72B teacher (87.1%)—while...