"Metis: Memory Foundation Model" Most AI agents still use memory as an external RAG-style module, so the model retrieves old text instead of actually remembering. This paper makes memory native to the Transformer. So past interactions are compressed into dynamic layer states and read through memory attention during normal forward passes. The model weights stay frozen at inference, but its memory state updates without gradients, giving the model persistent memory inside the backbone. Still early and lossy, but this is yet another paper with a big step toward agents that remember natively instead of outsourcing memory to a database.

“AREX: Towards a Recursively Self-Improving Agent for Deep Research” Deep research agents often fail not because they need more search, but because they don’t know which parts of an answer are already verified and which constraints are still unresolved. This paper turns research into a recursive loop: answer, verify constraint-by-constraint, preserve evidence, then refine only the weak parts. AREX adds learned context updates and step-aware RL, achieving extremely strong results with 4B and 122B-A10B models across various agentic benchmarks.

Kimi K3 + Tinker is a really nice setup for an autoresearch loop With most of the training plumbing off the table, Kimi’s code changes stay focused on the experiment itself. This makes it easier to follow what changed and steer the next runs. Reproducing the Self-Distilled RLVR paper, Kimi one-shots the baseline implementation, ran 19 experiments across six configurations, and wrote the final report (plus a bonus, unsolicited Chinese version)

GPT 5.6 Sol can one-shot convert an arXiv paper into an interactive Marimo notebook! Great for papers best understood hands on (lots of fun examples in interpretability, inference engineering, agent harnesses, benchmarking, and more) Play around with the notebook, inspect the code, or try the same workflow with your own agents below

is one of the coolest projects I've seen. It automatically reproduces results in a paper using agents that automatically spin up GPUs on cloud. I've just sent my most recent paper and it automatically reproduced it end to end 🙂 @askalphaxiv

"LLM-as-a-Verifier: A General-Purpose Verification Framework" The key idea of this paper is that it does not ask for one rough score, it reads the model’s full uncertainty over scores, which helps to make the judgment much more fine-grained. This approach lets agents pick better solutions, track progress, and learn from denser feedback.

Introducing the Global Researcher Map 🌎 We mapped every AI researcher into a visual landscape you can explore Search your favorite authors, topics, or institutions, and see who’s behind the work

"LLM-as-a-Verifier: A General-Purpose Verification Framework" The key idea of this paper is that it does not ask for one rough score, it reads the model’s full uncertainty over scores, which helps to make the judgment much more fine-grained. This approach lets agents pick better solutions, track progress, and learn from denser feedback.

Introducing the Global Researcher Map 🌎 We mapped every AI researcher into a visual landscape you can explore Search your favorite authors, topics, or institutions, and see who’s behind the work

Sakana Fugu Technical Report Instead of training one larger model, Sakana AI trains an orchestrator that reads each query and dynamically routes or composes GPT-5.5, Gemini-3.1-Pro, Claude Opus 4.8 and other agents into query-specific workflows. With Fugu being the fast router, and Fugu-Ultra being the deep multi-agent conductor, trained with SFT, evolutionary strategies and GRPO to build adaptive scaffolds. The idea is to have the model pick GPT for math, Gemini for science and recall, Opus for debugging, then synthesize them when no single agent is best. This router is able to get SoTA results across SWE-Bench Pro, Terminal Bench, LiveCodeBench, GPQA-Diamond, CharXiv and more, demonstrating the potential of orchestration being a practical alternative beyond training.

Here’s a fun comparison between GLM 5.2 and Opus 4.8 on a one-shot reproduction of the SDPO paper This is a hard task: the model must resolve messy verl issues and then run ablations to completion and confirm the paper’s claims. - GLM 5.2 costs $6.21 while Opus 4.8 cost us $46.35 - Both models spent a bulk of their tokens resolving initial verl issues. GLM 5.2 attempted 14 failed runs before first success while Opus 4.8 attempted 9 runs. - GLM 5.2 surprisingly took 2.65M tokens (excl re-reads) compared to 4.53M tokens for Opus 4.8

What if you have transformers learn to think in both latent states & tokens? This Microsoft paper adds a self-supervised next-latent prediction objective to standard next-token training, where a lightweight dynamics model learns to predict the transformer’s next hidden state from the current hidden state and next token. This pushes transformer representations toward compact belief states with coherent transition dynamics, improving world modeling, reasoning, planning, and enabling variable-length self-speculative decoding with up to 3.3x inference speedup.

OCR but with a working memory...?! This paper, Unlimited OCR, replaces decoder full self-attention with Reference Sliding Window Attention, giving the model a working memory where each token attends to the fixed visual and prompt references plus only the most recent output tokens, making decode KV cache constant instead of linear in generation length. Combined with DeepEncoder’s 16x visual compression, this working-memory-style attention enables one-shot multi-page OCR with stable memory and latency, while improving OmniDocBench performance over DeepSeek OCR.

Introducing GLM 5.2 for autoresearch GLM 5.2 is the first open weights model we've tried on our autoresearch pipeline that's proven capable for real research tasks. With Fable 5's restrictions on research, having an open weights alternative is a huge win for open source Watch it carry out fully async vs colocated sync RL training on Harbor code contests across two 8xH100 nodes on top of SkyRL. Resolves setup issues, tracks runs to completion, and produces a full comparison of throughput and reward stability

Introducing autoresearch for arXiv papers Change 'arxiv' to 'autoarxiv' in any paper URL An agent deploys to resolve setup issues on the codebase, run a minimal reproduction, and estimate full replication cost. Read more below