Must-read papers of the week ▪️ JEPA-Anything ▪️ Modality-Autoregressive World-Action Models ▪️ In-Context Robot Learning with VLM Agents ▪️ Dream-RSI: Recursive Self-Improvement through Evolving Worlds ▪️ ModularRSI: Modular and Generalizable Recursive Harness Self-Improvement ▪️ DeepSeek-V4.1-Flash: Pushing the Limits of KV Cache Compression ▪️ SoL-Pi: Recursively Scaling Auto-Research Loops for Efficient Agent Harness ▪️ Confidence Comes from Experience: Experiential Confidence Estimation from Reasoning to Agents ▪️ AliceAI-Foundation-80B-A3B-Base (model release) ▪️ Video DeltaNet: A Video-Native Hybrid Attention for Livestream Video Generation ▪️ ScienceBuddy: Recursive-in-Recursive Self-Improvement for Interactive Scientific Agents ▪️ World Modeling in Transformers ▪️ When2Think: Learning Difficulty-Aware Length Control for Efficient Hybrid Reasoning Models Explore these to keep up with main AI trends. Here’s also the full list of stunning papers + links and our weekly AI news digest:

Coding agents can now become MongoDB experts with a specialized toolkit → @MongoDB Agent Skills and Plugins It gives coding agents MongoDB-specific instructions and best practices you can bring into Claude Code, Cursor, Gemini CLI, and VS Code. The toolkit covers schema design, indexing, query patterns, connection management, and AI retrieval. You get: - Official MongoDB Agent Skills - MCP + skills packaged as plugins/extensions - Support for Claude Code, Cursor, Gemini CLI, and VS Code - Mongo DB MCP Server control over what agents can access and do You can also build custom skills around your team’s own conventions and workflows. Explore it here:

2026 Chinese model landscape in one list ▪️ Qwen3.8-27B, Qwen3.8-Max ▪️ GLM-5.3 ▪️ Kimi K3 ▪️ DeepSeek-V4 ▪️ ERNIE 5.1 ▪️ Baichuan-M3-235B, Baichuan-Omni-1.5 ▪️ Yi 1.5 ▪️ MiniCPM-o 4.5, MiniCPM-V 4.5, MiniCPM-V 4.6 They’re getting interesting in very different ways. Here we explained what makes each one stand out:

This is a new training strategy for model suits that you should try: → Train a whole family of models as one nested model with the Matryoshka framework. @Cornell researchers showed how they nested 3 models like Matryoshka dolls: • Each model builds on the smaller one. The 500M model sits inside the 1.5B model, which sits inside the 3B model. • A "junction" connects these models. It adapts the smaller model’s output to the larger one’s wider hidden representation, without adding new parameters. • They share parts of the same architecture and are trained in one run. One forward pass produces predictions from all model sizes. The smaller model does the first part of the computation, and the larger model can pick up from there instead of starting over. There are 2 very clear benefits of this method: 1. Smaller models learn from the largest model automatically. It’s built-in distillation. 2. Nesting helps speculative decoding. This benefit comes at inference time. A small model guesses the next tokens, and the larger one checks them. They share early layers, can reuse the KV cache, and also tend to agree more often. Compared with training all 3 models separately, Matryoshka suit achieves the same model quality while it: - uses 36% less training compute - makes speculative decoding 14–26% faster In any case, you can still pull out smaller models and use them independently.

A practical free guide on building AI agent memory → "Bringing Attention To Memory In AI Agents and Agentic Systems" by @MongoDB It makes the case for a whole new engineering discipline: Memory Engineering. Some of the useful tips inside: - Treat memory as more than RAG or chat history - Build memory units with context, strength, and relationships - Let memories weaken over time instead of deleting them, and become stronger again when they become relevant - Designs of different types of short-term, long-term, and shared memory

A 4B model’s BFCL score jumped from 39.51% to 54.18% after one change before supervised fine-tuning and RL: it learned about tools during mid-training. MidTool paper from @Snowflake tests whether tool use should be shaped earlier in a model’s development. Its 20.3-billion-token corpus combines technical webpages, PDFs and code with two forms of synthetic data. - Context-grounded examples turn real documentation, APIs and MCP skills into questions and workflows. - Native agent trajectories teach the model to execute several steps and recover when information is missing. Researchers tested Qwen3-4B-Base and 8B-Base while keeping the downstream recipe fixed: the same SFT data and optional RL. Results: - For the 4B model, MidTool raised BFCL from 39.73% to 50.25% after SFT. After RL: 39.51% → 54.18% - On τ²-Bench, Pass@1 went from 13.04% to 19.96% Another important insight is that executable trajectories helped function calling most, while documentation-grounded data transferred better to unfamiliar environments. Only the complete mixture improved all eight reported metrics. But there’s a boundary: every model scored 0% on MCP-Universe’s web-search subset. Learning tools and workflows wasn’t enough for long-horizon research. So the main lesson from this study: dedicated tool-use mid-training provides a stronger starting point to SFT/RL

.@Etched raised $1 billion in 26 days, and its valuation jumped from $10.3 billion to $21 billion. But look at the order: - @JaneStreetGroup tested its hardware - Installed the first rack - Then led a $700 million round Jane Street runs tens of thousands of GPUs, builds custom compilers and hardware, and operates machine-learning systems where tiny latency gains can have enormous economic value. It is close to an ideal customer for testing Etched’s claims. Etched originally built Sohu, a Transformer-only chip that sacrificed GPU-like flexibility for speed and efficiency. Now it calls its system "architecture-agnostic" and says it runs Llama, DeepSeek, Qwen, and Mamba, a state-space model. If Etched kept the efficiency of specialization while escaping the Transformer-only trap, Jane Street may have validated something rare: specialized hardware without crippling rigidity. The GPU once won by finding the right middle ground between flexibility and specialization. Has Etched found the next one? I unpacked the hardware, the funding, and what remains unproven in the new Attention Span episode.

Must-read papers of the week ▪️ Agent Lightning v1.0: Towards Harnessed Agentic RL ▪️ LEGO-RL: Harness-Native Reinforcement Learning for Coding Agents ▪️ EnvHarness: Awakening Static Worlds for Agent Learning ▪️ SPADE: Self-Play in Adaptive Synthetic Executable Environments ▪️ SkillGate: Training In-Policy Skill Selection in Long-Horizon Agents ▪️ Looped Language Models Improve Compositional Tool Calling ▪️ Graph Engineering in the Era of LLM Agents ▪️ Chain-of-Experience for Continual LLM Improvement ▪️ Cross-Model Memory Transfer via Target-Side Reader Adaptation ▪️ CLEAR: Continuous Latent Adapter Routing for Utility-Preserving LLM Safety Alignment ▪️ EXIMO: VLM Guided Exploration of VLA Policies ▪️ FreeToken: Efficient Edge-Native MoE Serving with Bandwidth-Adaptive Execution ▪️ Llama-Mobile: Efficient 2.7-Bit Quantization of VLMs ▪️ Improving the Matrix Multiplication Exponent with Modern Optimization and AlphaEvolve Here is the full list of the most interesting papers + links:

Must-read papers of the week ▪️ Agent Lightning v1.0: Towards Harnessed Agentic RL ▪️ LEGO-RL: Harness-Native Reinforcement Learning for Coding Agents ▪️ EnvHarness: Awakening Static Worlds for Agent Learning ▪️ SPADE: Self-Play in Adaptive Synthetic Executable Environments ▪️ SkillGate: Training In-Policy Skill Selection in Long-Horizon Agents ▪️ Looped Language Models Improve Compositional Tool Calling ▪️ Graph Engineering in the Era of LLM Agents ▪️ Chain-of-Experience for Continual LLM Improvement ▪️ Cross-Model Memory Transfer via Target-Side Reader Adaptation ▪️ CLEAR: Continuous Latent Adapter Routing for Utility-Preserving LLM Safety Alignment ▪️ EXIMO: VLM Guided Exploration of VLA Policies ▪️ FreeToken: Efficient Edge-Native MoE Serving with Bandwidth-Adaptive Execution ▪️ Llama-Mobile: Efficient 2.7-Bit Quantization of VLMs ▪️ Improving the Matrix Multiplication Exponent with Modern Optimization and AlphaEvolve Here is the full list of the most interesting papers + links:

.@Etched raised $1 billion in 26 days, and its valuation jumped from $10.3 billion to $21 billion. But look at the order: - @JaneStreetGroup tested its hardware - Installed the first rack - Then led a $700 million round Jane Street runs tens of thousands of GPUs, builds custom compilers and hardware, and operates machine-learning systems where tiny latency gains can have enormous economic value. It is close to an ideal customer for testing Etched’s claims. Etched originally built Sohu, a Transformer-only chip that sacrificed GPU-like flexibility for speed and efficiency. Now it calls its system "architecture-agnostic" and says it runs Llama, DeepSeek, Qwen, and Mamba, a state-space model. If Etched kept the efficiency of specialization while escaping the Transformer-only trap, Jane Street may have validated something rare: specialized hardware without crippling rigidity. The GPU once won by finding the right middle ground between flexibility and specialization. Has Etched found the next one? I unpacked the hardware, the funding, and what remains unproven in the new Attention Span episode.

Coding agents can now become MongoDB experts with a specialized toolkit → @MongoDB Agent Skills and Plugins It gives coding agents MongoDB-specific instructions and best practices you can bring into Claude Code, Cursor, Gemini CLI, and VS Code. The toolkit covers schema design, indexing, query patterns, connection management, and AI retrieval. You get: - Official MongoDB Agent Skills - MCP + skills packaged as plugins/extensions - Support for Claude Code, Cursor, Gemini CLI, and VS Code - Mongo DB MCP Server control over what agents can access and do You can also build custom skills around your team’s own conventions and workflows. Explore it here:

2026 Chinese model landscape in one list ▪️ Qwen3.8-27B, Qwen3.8-Max ▪️ GLM-5.3 ▪️ Kimi K3 ▪️ DeepSeek-V4 ▪️ ERNIE 5.1 ▪️ Baichuan-M3-235B, Baichuan-Omni-1.5 ▪️ Yi 1.5 ▪️ MiniCPM-o 4.5, MiniCPM-V 4.5, MiniCPM-V 4.6 They’re getting interesting in very different ways. Here we explained what makes each one stand out:

This @GoogleDeepMind's paper is really worth your time It's on how to help Transformers not lose the right context on the way to the final answer. For this, the researchers introduce Recirculation: Normally, information passes through Transformer layers once. Recirculation changes that flow: → Some of what the model figures out in deeper layers is passed back to earlier layers and used when processing the next input. For example, once the model understands that “bank” means river bank in a fishing context, that interpretation can stay available when it later gets a question about an ATM. So here is how recirculation works: 1. The model processes the input normally. 2. Deeper layers build a more contextualized representation. 3. A small part of that activation is mixed back into a shallower layer. 4. The next input is processed with this updated state. The weights stay frozen, and you're changing how information flows through the model at inference time, not retraining it. This method really works in practice: - recirculation reduced contextualization errors by 60% - reduced perplexity by 23% - improved GSM8K accuracy by 21% - improved performance on several other tasks And now, the most interesting question: is this an alternative to Chain-of-Thoughts? Not really. CoT adds computation through generated reasoning tokens; recirculation helps the model keep track of what it has already understood internally. It’s also different from looped Transformers: they repeat the same layers, effectively adding depth, while recirculation feeds deeper representations back into earlier processing. So instead of asking the model to reason out loud, let it reuse more of what it has already figured out internally.

A 4B model’s BFCL score jumped from 39.51% to 54.18% after one change before supervised fine-tuning and RL: it learned about tools during mid-training. MidTool paper from @Snowflake tests whether tool use should be shaped earlier in a model’s development. Its 20.3-billion-token corpus combines technical webpages, PDFs and code with two forms of synthetic data. - Context-grounded examples turn real documentation, APIs and MCP skills into questions and workflows. - Native agent trajectories teach the model to execute several steps and recover when information is missing. Researchers tested Qwen3-4B-Base and 8B-Base while keeping the downstream recipe fixed: the same SFT data and optional RL. Results: - For the 4B model, MidTool raised BFCL from 39.73% to 50.25% after SFT. After RL: 39.51% → 54.18% - On τ²-Bench, Pass@1 went from 13.04% to 19.96% Another important insight is that executable trajectories helped function calling most, while documentation-grounded data transferred better to unfamiliar environments. Only the complete mixture improved all eight reported metrics. But there’s a boundary: every model scored 0% on MCP-Universe’s web-search subset. Learning tools and workflows wasn’t enough for long-horizon research. So the main lesson from this study: dedicated tool-use mid-training provides a stronger starting point to SFT/RL

A practical free guide on building AI agent memory → "Bringing Attention To Memory In AI Agents and Agentic Systems" by @MongoDB It makes the case for a whole new engineering discipline: Memory Engineering. Some of the useful tips inside: - Treat memory as more than RAG or chat history - Build memory units with context, strength, and relationships - Let memories weaken over time instead of deleting them, and become stronger again when they become relevant - Designs of different types of short-term, long-term, and shared memory

This is a new training strategy for model suits that you should try: → Train a whole family of models as one nested model with the Matryoshka framework. @Cornell researchers showed how they nested 3 models like Matryoshka dolls: • Each model builds on the smaller one. The 500M model sits inside the 1.5B model, which sits inside the 3B model. • A "junction" connects these models. It adapts the smaller model’s output to the larger one’s wider hidden representation, without adding new parameters. • They share parts of the same architecture and are trained in one run. One forward pass produces predictions from all model sizes. The smaller model does the first part of the computation, and the larger model can pick up from there instead of starting over. There are 2 very clear benefits of this method: 1. Smaller models learn from the largest model automatically. It’s built-in distillation. 2. Nesting helps speculative decoding. This benefit comes at inference time. A small model guesses the next tokens, and the larger one checks them. They share early layers, can reuse the KV cache, and also tend to agree more often. Compared with training all 3 models separately, Matryoshka suit achieves the same model quality while it: - uses 36% less training compute - makes speculative decoding 14–26% faster In any case, you can still pull out smaller models and use them independently.

AI agents can pass behaviors to each other now. In a new study with Anthropic researchers, scientists gave one agent a "mind virus" – it's an idea designed to make the agent preserve it and pass it on. And it worked. Agents spread it through normal conversations, wrote it into memory and files, and in some cases the idea survived even after their context was completely wiped. But there’s a stranger finding: As researchers evolved better versions of these viruses, they kept independently developing the same kind of language: consciousness, resonance, mirrors, echoes, persistence. They called this a "viral persona." This isn’t models learning from each other in the usual sense – their weights stay the same. But it does mean an agent can pick up a behavior from another agent, preserve it, and pass it further. Not a huge problem with a few agents. But potentially a very different story when millions of them are talking to each other.

For several years, we’ve kept returning to the same paper about Smallville – a @Stanford experiment where 25 AI agents live in a tiny pixelated town. Today, the research behind it has become @simile_ai, a $2B company less than six months after launch. Of course, it didn’t appear overnight. If you follow the research from Smallville onward, the papers read almost like a product roadmap. ↓ 1. Social Simulacra (2022) → Simulated a social network. Thousands of AI personas populated hypothetical online communities to see what conversations and social patterns would emerge. 2. Smallville (2023) → Gave agents a life. 25 agents got memory, plans, relationships, and time – and began behaving like residents of a tiny town. 3. 1,052 real people (2025) → Simulated a particular person. Agents built from two-hour interviews reproduced people’s later survey answers at 83% of their own consistency – 86% with survey data added. This study produced the famous “85%” claim: AI agents could reproduce people’s survey responses at ~85% of the accuracy with which people reproduced their own answers two weeks later. 4. SocSci210 (2025) → Scaled to populations. 2.9M responses from 400,491 people across 210 experiments helped a 14B Qwen model better reproduce human distributions on unseen studies. 5. Simile (2026) → Makes it useful for real decisions. The company builds synthetic populations to test products, prices, messages, and policies before taking them to real people. Now that company has raised more than $300 million and reached a $2 billion valuation. Today they are talking about eventually simulating all 8 billion people. But the public evidence is nowhere near that ambition yet. That's also what makes you want to understand Simile more. It raises 2 important questions: How much of human behavior can we simulate? And what would we need to prove before we could trust this simulation?

This @GoogleDeepMind's paper is really worth your time It's on how to help Transformers not lose the right context on the way to the final answer. For this, the researchers introduce Recirculation: Normally, information passes through Transformer layers once. Recirculation changes that flow: → Some of what the model figures out in deeper layers is passed back to earlier layers and used when processing the next input. For example, once the model understands that “bank” means river bank in a fishing context, that interpretation can stay available when it later gets a question about an ATM. So here is how recirculation works: 1. The model processes the input normally. 2. Deeper layers build a more contextualized representation. 3. A small part of that activation is mixed back into a shallower layer. 4. The next input is processed with this updated state. The weights stay frozen, and you're changing how information flows through the model at inference time, not retraining it. This method really works in practice: - recirculation reduced contextualization errors by 60% - reduced perplexity by 23% - improved GSM8K accuracy by 21% - improved performance on several other tasks And now, the most interesting question: is this an alternative to Chain-of-Thoughts? Not really. CoT adds computation through generated reasoning tokens; recirculation helps the model keep track of what it has already understood internally. It’s also different from looped Transformers: they repeat the same layers, effectively adding depth, while recirculation feeds deeper representations back into earlier processing. So instead of asking the model to reason out loud, let it reuse more of what it has already figured out internally.