Banger paper from Meta Superintelligence Labs. They find something super interesting and unexpected. (bookmark it) Base models with a light harness often solve more agentic tasks than their RL post-trained versions when both get enough samples. Post-trained models win on
@dair_ai
-
NVIDIA paper on long-running AI agents and context loss
By
–
Banger paper from NVIDIA on long running agents. A model can accept 128K tokens of context and still make more mistakes the longer it works through a task. If your agent loses its place partway through a long table or ledger, this work measures what causes it. The setup:
-
Microsoft paper on compressing agent context at test time
By
–
New paper from Microsoft on compressing agent context at test time. If your agent gets worse as its interaction history grows, this method is worth a look. FOCUS asks which past interactions the agent's next decisions actually depend on. It keeps those units of the history and
-
Meta paper on controlling long agent runs with benchmarks
By
–
Super interesting paper from Meta Superintelligence Labs on controlling long agent runs. They use the same workers and same budget, and ProgramBench goes from 63.7% to 71.5% with GPT-5.5 when a dedicated controller decides what work to run next. Codex scores 58.0%. Here is how
-
Meta paper on agent harness optimization via branched search
By
–
Great paper from Meta on agent harness optimization. Meta-Harness-style search uses one development set and one proposal policy, so every edit follows a single path and can get stuck in a local optimum. This work splits the search into branches. Each branch keeps the
-
Meta’s Context Language Models manage their own context
By
–
Great paper for the weekend ahead. Meta presents Context Language Models, which can natively manage their own context.
-
Combining LLMs: analysis of routing, voting, cascades
By
–
When does combining LLMs help? Great analysis on combining language models, measured across 67 models from 21 providers. Any policy that routes, votes, cascades, or runs a mixture of agents and then returns one model's answer is bounded above by 1 minus beta, where beta is the
-
NVIDIA SOLAR automates speed-of-light performance analysis from PyTorch/JAX
By
–
NEW paper from NVIDIA. (bookmark it) Speed-of-light performance analysis tells you the theoretical floor of a workload, but teams still derive it by hand and freeze it. SOLAR automates the whole thing straight from PyTorch or JAX source. An LLM frontend translates arbitrary
-
Agent memory becomes data management system for LLMs
By
–
// Agent memory is a data system now // Great paper on long-term memory for LLM agents. (bookmark it) Agent memory has grown from simple retrieval into a full data-management layer with storage, retrieval, update, consolidation, and lifecycle governance. Yet most evaluations
-
Why Model Routers Fail for Coding: Information Deficit
By
–
What is actually limiting model routers for coding tasks? Most routers treat picking a model as a static, one-off classification. This paper identifies the real bottleneck as information deficit. Simply augmenting a vanilla LLM router with task-dimension-level performance