Model Releases

openbmb/MiniCPM-RobotManip — Vision-language-action model for robotic manipulation built on MiniCPM. If you’re working on embodied AI and need something that can actually do things with a robot arm rather than just describe them, this is worth a look. Custom robotics code baked in.

openbmb/MiniCPM-RobotTrack — Companion visual tracking model for the same stack. Handles perception side of embodied AI — object tracking, feature extraction for downstream manipulation. Same architecture, different head.

Open Source Releases

hound-mcp 11.1.0 — MCP server that gives agents web research chops with anti-bot bypass and graceful HTTP fallback. Uses primp + patchright directly (no scrapling), token-optimized tool defs, cross-platform CLI, self-update with rollback. Practical if you’re building agents that need to actually browse the modern web without getting blocked.

ternary-QAT 0.1.0 — Ternary quantization-aware training for transformers. Pushes weights to 1.58-bit (ternary) while preserving accuracy via QAT fine-tuning. From the Ternary-Bonsai project. Extreme compression territory — useful when you need models to fit on edge devices or just want to slash inference costs.

transcribe.cpp — GGML-based STT in C++ supporting 16+ model families (Whisper, Moonshine, Distil-Whisper, etc.). CPU-optimized, quantized, minimal deps. If you need offline speech-to-text without dragging in PyTorch, this is the lightweight alternative to whisper.cpp.

Research Worth Reading

Deterministic Replay for AI Agent Systems — Tackles the nightmare of non-determinism in LLM agents (sampling variance, API state drift, environment noise). Proposes a replay framework for faithful re-execution — critical for debugging and auditing. Finally, a way to answer “why did the agent do that?” with something better than “vibes.”

Rater State Bias in RLHF Preference Data — Finds that pairwise preferences in RLHF reflect rater fatigue/stress over time, not just output quality. Structured confound that biases your reward model. Audit framework included to detect it in your datasets. Your “human feedback” has human problems.

PlanFlip: Attacking Multi-Agent LLM Systems via Planning-Phase Prompt Injection — Single prompt injection into the Planner cascades through all downstream Executor and Critic sub-tasks. Planning phase = critical attack surface. Multi-agent systems inherit single-point-of-failure vulnerabilities. 🤖

Masked Diffusion Language Models as World Models for Agentic RL — Masked diffusion LMs work as steerable world models for agentic RL. Enables diverse environment generation and solves sparse-reward mode collapse in long-horizon tasks. Diffusion isn’t just for images anymore.

PPO-HSC: Wide-Area Policy Coverage Optimization — Addresses mode collapse in LLM fine-tuning (the “Invisible Shackles” of standard RLVR) via high-order sampling coverage. Exploratory RL framework that actually explores instead of collapsing to a few high-reward modes.

ColGraphRAG: Late-Interaction Evidence Retrieval for Multimodal GraphRAG — Late-interaction retrieval preserves patch/token-level structure in graph-linked images that bi-encoders flatten away. Better fine-grained evidence ranking for multimodal GraphRAG. Bi-encoders were throwing away signal.

AI Dev Tools

Moonshine — Ultra-low latency voice stack: STT, intent recognition, TTS all streaming, on-device capable. Built for real-time voice agents where every millisecond counts. If you’re building voice interfaces and latency is killing UX, this is the stack to benchmark against.

academic-research-skills — End-to-end academic pipeline as agent skills: lit search → paper writing → peer review sim → revision → finalization. Citation management, methodology validation, reproducibility checks. For Claude Code. Automates the tedious parts of research so you can focus on the ideas (or the coffee). ☕

Today’s Synthesis

If you’re building multi-agent systems, PlanFlip just handed you a threat model: a single prompt injection in the planning phase cascades through every downstream executor and critic. The planner is your new critical attack surface. But detecting that cascade in production requires something most agent frameworks lack — Deterministic Replay . Non-determinism from sampling variance, API drift, and environment noise means you can’t reliably re-run an attack to understand it. The replay framework lets you capture full execution traces (including planner outputs) and re-execute faithfully for forensics. Pair these: instrument your planner with structured logging, feed traces into a replay store, and you get an audit trail that answers “why did the agent do that?” with evidence instead of vibes. Bonus: the same infrastructure catches PPO-HSC ’s mode collapse during RL fine-tuning — replay lets you compare exploration trajectories across runs to verify you’re actually covering policy space, not collapsing to a few high-reward modes. Build replay first; debug everything else after. 🤖