CCTest · Blog
When a Model Knows but Refuses: PIR Probes Hidden Recognition
Evaluation & Benchmarks
cctest.ai

When a Model Knows but Refuses: PIR Probes Hidden Recognition

A new study introduces PIR, a method that examines a language model’s internal states to distinguish missing knowledge from knowledge that is being concealed. The approach continues to detect recognized answers under deception, sandbagging, password locks, and circuit-based restrictions.

Read more
CCTest · Blog
From One to Many and Back: Category-Aware Training for SWE Agents
Coding AI
cctest.ai
Coding AI

From One to Many and Back: Category-Aware Training for SWE Agents

A new study addresses the uneven progress that can emerge when heterogeneous software-engineering tasks are trained in a shared reinforcement-learning pool. It develops category experts with an iterative Refresh–Repair–Expand loop, then consolidates them into one deployable model through label-routed multi-teacher on-policy distillation.

Read more
CCTest · Blog
ShieldVLA Teaches Vision-Language-Action Models When to Recover
Robotics & Physical AI
cctest.ai

ShieldVLA Teaches Vision-Language-Action Models When to Recover

ShieldVLA brings Hamilton-Jacobi reachability into VLA safety fine-tuning, using visual observations to estimate whether a robot can still recover safely. The paper reports a 57% average reduction in cumulative safety cost across five navigation and manipulation benchmarks, alongside a 0.13 success-rate gain over SafeVLA.

Read more
CCTest · Blog
vLLM Adds Hardware-Agnostic Layers to Balance Frontier Speed and Portability
Inference & Serving
cctest.ai

vLLM Adds Hardware-Agnostic Layers to Balance Frontier Speed and Portability

vLLM is moving toward flat, hardware-specific model implementations for frontier performance while introducing hardware-agnostic layers for older GPUs, external accelerators, and broader model support. On H100, the new approach is within 3.4% of the native implementation in total token throughput.

Read more
CCTest · Blog
Grounded Action Models Put 3D Object Grounding at the Core of Robot Control
Robotics & Physical AI
cctest.ai

Grounded Action Models Put 3D Object Grounding at the Core of Robot Control

Grounded Action Models (GAMs) make metric 3D grounding an explicit part of robot action prediction. By turning language, point, and box prompts into a shared object-centric representation, GAMs improve robustness when targets move, scenes change, or tasks require long-horizon control.

Read more