KV-PRM reads the LLM's KV cache instead of re‑encoding whole trajectories, turning reward scoring from quadratic to linear time. The trick slashes FLOPs by up to 5,000×, cuts latency 37× and memory 34×, while matching or beating text‑based PRMs on MATH, GSM8K and AIME.
Zhipu AI co‑founder Tang Jie used an internal staff letter to unveil a two‑year “Touch High” plan that puts foundation‑model research, open‑source releases like GLM‑5.2, and billions in safety work ahead of short‑term product revenue. The move signals China’s biggest AI lab is betting on AGI breakthroughs that stay publicly accessible, reshaping global competition.
The paper introduces a shared selective persistent memory that captures reusable context, task specs, schemas, tool configs, and constraints, while discarding session‑specific traces. In enterprise tests it lifts task completion from 79% to 96% and slashes token use by up to 97×, enabling agents to collaborate without re‑specifying everything.
The paper builds transformers from bounded, named operators that act like fuzzy set detectors, then forces them to stay crisp during training with a variance-floor loss. The resulting models make 78% of feed‑forward units and half of attention channels interpretable, and edits become 50‑184× more localized, enabling surgical manipulation of concepts.
The paper derives exact solutions for how linear concept directions form during training, showing that data and target geometry, depth, and initialization scale dictate the final abstraction. It extends the theory to nonlinear nets, revealing that ReLU and erf activations attenuate abstraction differently, and validates the findings on models like DINOv3 and Gemma 4, improving linear probe generalization in LLMs.
ARCANA coordinates a set of specialized agents to crack ARC‑AGI‑2 puzzles within tight inference budgets. By iteratively grounding raw grids, proposing DSL programs, executing them symbolically, and reflecting on failures, it beats prior methods on the benchmark, showing that structured multi‑turn program synthesis can scale under strict test‑time and hardware limits.
The authors introduce the Hypothesis Evolution Protocol, a framework that forces LLM agents to log hypothesis generation, testing, and belief updates as explicit, auditable steps. In materials‑science benchmarks, the protocol improves transparency and enables humans to verify and build on AI‑driven scientific reasoning.
Modern AI excels at reasoning, coding, and tool use but stays within a preset conceptual frame. This limits open‑ended innovation because models can’t create new representational primitives or reliably verify their value. The paper defines vocabulary and verifier gaps and proposes research directions to give AI generative representation capabilities.
The LMSYS Chatbot Arena turns LLM evaluation into a blind, crowdsourced showdown, a “Pepsi Challenge” that pits anonymous models against each other. Using a Bradley‑Terry rating system, it surfaces real‑world helpfulness and separates even top models on coding, reasoning, and long‑context tasks, giving developers a clear view of which model truly outperforms the rest for their use‑case.
1X unveiled its Neo humanoid’s new tendon‑driven hands, packing 25 backdrivable joints that sense pressure and shear. The force‑transparent design lets the robot feel objects and adjust grip in real time, opening up tasks from pouring tea to handling fragile items. This hardware leap removes the long‑standing dexterity ceiling for humanoid platforms.
The post argues that AI safety’s biggest obstacle is a lack of political will, not missing research. It quantifies how shifting from today’s "Plan D" to a coordinated international "Plan A" could slash takeover risk from ~45% to ~7%, making collective advocacy the first urgent step.
Apple sued OpenAI, alleging the AI firm stole confidential designs and manufacturing details to jump‑start its own consumer‑hardware line. The complaint says OpenAI recruited Apple staff, coaxed candidates to bring proprietary components, and leveraged over 400 ex‑employees to melt Apple’s trade secrets. The case could shape who controls the next wave of AI‑driven devices.
Researchers show that LLM refusal stems from a few activation directions and introduce activation‑guided suffix attacks that bypass safety layers. The method speeds up jailbreaks 33× and reveals that safety signals are distributed across the network, challenging current alignment approaches.
A Google‑Sheets based Epistemic Audit lets AI‑safety researchers map, quantify, and track belief uncertainties across key risk domains. By turning vague gut feelings into measurable data, it highlights where confidence is slipping and points to the questions that could shift risk assessments the most.
You’re caught up.
That was all 14 stories in the 13 July issue.