The Agent Teaches Itself With Skills It Wrote: How SEED Fixes the Sparse-Reward Problem
The real bottleneck in agentic RL is that the reward arrives only once, at the very end of a trajectory. SEED turns that single sparse signal into dense per-...
Research
Longer Thinking, Linear Cost: How the Markovian Thinker and Delethink Redesign Long Reasoning
The real cost of long reasoning comes from the state growing without bound. Markovian Thinking breaks reasoning into fixed-size chunks and passes only a shor...
Why Dynamic Batch Tuning Failed to Beat Static Scheduling: An Honest Failure Analysis of a Multi-Tenant vLLM Serving Control Loop
Does an agent that observes GPU telemetry and adjusts batch size and concurrency in real time outperform static Kueue admission control? The simulation said ...
Harness Design and Post-Training Aren’t Separable: Harness-Aware Post-Training and LLM Agent Performance
Tool-using LLM agents run on top of a harness that wraps the model. A recent arXiv paper shows empirically that treating harness design and post-training as ...
Putting a Safety Net Under Autonomous Skill Rollout with Canary Gates: Trace-Based Regression Detection and Automatic Rollback
How should an agent harness that rewrites its own code overnight safely roll out the result the next day? An experiment pairing canary releases with automati...
Controlling Reasoning Effort - How LLMs Learn Low, Medium, and High Thinking Modes
With GPT-5.6 shipping five or six reasoning effort settings per size, effort control has become table stakes for reasoning models. We dig into what training ...
Skill Graveyard: Safely Deprecating Skills in an Agent Harness With Over 2,000 Units
If your team has kept expanding its catalog of agent skills or MCP tools, you’ve probably felt routing accuracy start to slip at some point. What happens if,...
Autonomous Remediation or Human Escalation: Calculating the Safety Boundary for Kubernetes GPU Incident Response
Any team trying to hand Kubernetes GPU incident remediation to an LLM agent eventually runs into this question: how long should the agent be left to fix thin...
If You Want an LLM to Generate Screens From a Design System, Build the Gate Before the Model
We expected fine-tuning to win. The data said otherwise. Here’s how we ended up with a generator that respects TDS without any training, and the next problem...
When Can You Trust the Cheap Model: Cutting Multilingual Document OCR Costs with a Confidence-Gated VLM Cascade
Even in an era where sub-billion-parameter, ultra-lightweight VLMs read documents well, the truly hard documents still need a large model. We introduce an an...
Auditing an Autonomous Nightly Paper-Generation Pipeline Over 8 Days: A Case Study in Reliability, Diversity, and Research Integrity
This study audits a fully autonomous pipeline that generates a paper every night, not by inspecting the final output but by examining its actual operational ...
Validate Expensive, Then Descend: Cutting a Skill Fleet’s Model Cost with Code Gates
When a company’s recurring work is defined as skills, each skill can tell you, through measurement rather than intuition, exactly how far down the model tier...
Guard the Router and Quantization Hurts Less: An Accuracy-Cost Map for Selective NVFP4 Quantization in MoE
For engineers serving MoE models on H200 or Blackwell-class clusters, we introduce a paper that formalizes RASQ, an NVFP4 quantization policy that selectivel...
Less Verification Is Safer: Measuring the Cost-Quality Curve of Multi-Agent Verification Gates
The principle that every fan-out should close with adversarial verification is widely accepted, but how much verifier tier and how many skeptics to use has m...
Agentic RL Stops Waiting for the Group and Learns from One Rollout at a Time
When training long-horizon agentic tasks with RL, GRPO’s group sampling idles the GPU while it waits for the slowest rollout to finish. SAO, the method Tsing...
Remember the Decision, Not the Description: A Meta-Affiliated Study Reframes Agent Memory as a Rate-Distortion Problem
Long-running agents operate within limited memory, yet memory methods to date have organized the past using descriptive criteria such as relevance or summary...
LLM Research Ideas Lose on Range, Not Quality
Researchers at Yale and the University of Chicago compared human and LLM research ideas using 11,683 real papers. The conclusion is surprising. The problem w...
Orientation Before Execution: How GPT-5.6 Sol Broke ARC-AGI-3 for the First Time
ARC-AGI-3 measures whether an agent can figure out a situation and adapt on its own inside an interactive game with no instructions. GPT-5.6 Sol became the f...
Back from GRPO to PPO: How GLM-5.2 Stabilized RL with IcePop
The dominant trend in RL post-training these days is the GRPO family, which drops the critic. GLM-5.2, however, went back to PPO with a revived value model a...
The Night the Search Dictionary Fixed Itself: Closing a Skill-Routing Bottleneck with an Unattended Loop
After a prior diagnosis found that the real cause of routing failures in a large skill ecosystem was the retriever rather than decomposition, this study meas...
Neither the Retriever nor Decomposition: A Ceiling-Gap Diagnostic for Skill Routing Bottlenecks
For a harness that routes hundreds of skills, this diagnostic tells you with two numbers whether your next investment should go into decomposition or the ret...
Memory as an Action Space: AgeMem’s Unified Long- and Short-Term Memory Management
AgeMem folds long-term and short-term memory management into the agent’s policy, exposing store, retrieve, summarize, and discard as tool-based actions. We a...
The Hypothesis That Every AI Is Converging on the Same ‘Brain’: Reading the Platonic Representation Hypothesis
Vision models and language models trained on different data for different objectives are starting to represent data in the same way. MIT’s Platonic Represent...
Binding GPU TEE Attestation to Model Signatures: Attested Confidential Inference for Sovereign Multi-Tenant AI
Hospitals, banks, and government agencies running on-premises AI platforms must prove two things to regulators: that data never leaked outside the enclave, a...
An Evaluation That Trusts No Single Number: Building a Sovereign LLM-Judge Service on Binary Decomposition and Deterministic Gates
Using an LLM as a grader, the practice known as LLM-as-a-judge, is now the default in model development, but the evidence that piled up through 2026 shows th...
This Week’s Must-Read AI Papers (6/28-7/5): Metacognitive RL, Memory as a Skill, MCP Server Patterns
From dair.ai’s weekly paper roundup, we take a deep dive into three papers through a ThakiCloud lens: RLMF, which beats standard RL by up to 63% using metaco...
The Characters the Monitor Cannot Read: How Tokenizer Asymmetry Opens a Covert Channel Around AI Control
The core defense of AI control, the idea that an untrusted strong model can be controlled by a weaker, trusted monitor, rests on the assumption that the moni...
What Comes After AGI: DeepMind’s Four Paths to Superintelligence
Google DeepMind’s roughly 57-page report From AGI to ASI treats superintelligence not as a distant thought experiment but as a planning problem to prepare fo...
The Center of Gravity Has Shifted: Reading the Open Model Ecosystem Through the ATOM Report
The ATOM Report measures open language models across both downloads and inference usage in one place, and shows with data that Chinese open models overtook t...
Is Fine-Tuning Really Dead? A Survival Strategy Read from June 2026’s Verified Signals
As frontier LLMs and agent skills keep improving, the industry has started to feel that fine-tuning is no longer necessary. OpenAI is, in fact, winding down ...
ByteDance Dolphin Document Image Parsing: Fox Dataset and Benchmark Complete Analysis
A detailed analysis of ByteDance’s Dolphin project Fox dataset and benchmark, including the Analyze-then-Parse paradigm from ACL 2025 and a large-scale datas...