AI and infrastructure news
Filter by type: Paper51
Autonomous science limited by physical validation: allocating agents and validation bottlenecks
An arXiv paper argues that autonomous science is limited by physical validation capacity and proposes markets and institutions to allocate that scarce work.
HERMES: component-level harness with a resident LLM per code module
An arXiv paper proposes HERMES, where each repository component has a resident LLM that knows its code, with reported gains in migration and on four software engineering benchmarks.
Agent-R1: modular training for multi-turn LLM agents
The paper presents a reinforcement-learning framework for training LLM agents in multi-turn interactions with tools and sequential environments.
AutoSOTA automates iterative AI model improvement
Published on April 19, 2026, the post describes a system with eight specialized agents that reproduced and improved models, reporting 105 new state-of-the-art models across LLMs, computer vision, and time series.
Evolving coding-agent harnesses with evidence
A paper published on April 29, 2026 presents a framework for improving coding-agent harnesses. On Terminal-Bench 2, pass@1 rose from 69.7% to 77.0% over ten iterations.
Measuring performance when scaling agent systems
A December 11, 2025 article examines how coordination, model capability, and system and task factors influence agent performance.
Dual-play proposes training LLM reasoning with less supervision
Published on December 15, 2025, the paper presents dual-play: two models take specialized roles, one creating questions and the other solving them, to reduce reliance on external supervision.
Controlled evolution for algorithmic code optimization
A January 14, 2026 article presents an iterative code evolution method evaluated on EffiBench-X, using diverse strategies, feedback-guided mutations, and experience across tasks.
Layered memory for long-horizon research agents
A post dated April 19, 2026 describes ML-Master 2.0’s Hierarchical Cognitive Caching, with short-, medium-, and long-term memory, and reports a 56.44% medal rate on MLE-Bench after a 24-hour run.
MemRL learns to prioritize agent memories at runtime
Published on January 13, 2026, the description presents MemRL, an approach that learns the utility of episodic memories without changing language-model weights.
PARE-Bench tests proactive agents in stateful apps
Published on April 25, 2026, the paper presents 143 tasks for evaluating agents that infer goals, choose when to intervene, and coordinate actions across apps.
ReCode combines planning and action in LLM agents
Published on December 13, 2025, the article presents ReCode, a recursive code-generation approach that lets LLM agents switch between strategic planning and detailed actions.
StructMem proposes temporal hierarchical memory for agents
Published on April 24, 2026, the StructMem paper proposes hierarchical memory for long-term conversational agents, with event relationships intended to support temporal reasoning and multi-hop questions.
TACO proposes context compression for terminal agents
The paper describes a framework that learns context-compression rules from terminal-agent interactions and reports tests on three benchmarks.
RoPoLL proposes geometric median for LLM judge panels
Published on July 2, 2026, the paper formalizes LLM judge panels under a contamination model and introduces RoPoLL, tested with 13 judges and corruption rates up to 50%.
Autonomous research agents: applying the approach
Published on September 9, 2025, the paper describes reinforcement-learning-trained agents that search the web, navigate, use Python, and summarize prior results when context is limited.
163 procedures for scientific research agents
The article presents an open library of 163 procedures for research agents across 16 fields, including genomics, cheminformatics, medical imaging and study design.
Ablations show how to evaluate coding-agent harnesses
A paper published on September 26, 2026 reports 176 configurations tested on two benchmarks and describes differing effects from context summarization and explicit planning.
Agents can decide when to seek more context
An article published on August 22, 2026 models interactive agents’ choice between acting on assumptions and seeking context, accounting for the costs of those actions.
AREAL2.0 links real tasks to agent improvement
Published on July 10, 2026, the article presents a continuous-learning architecture that turns real tasks into training data and triggers policy updates.
Self-improvement in agents: making adaptation controllable
A survey published on July 16, 2026 examines modern agentic systems as foundation models integrated with prompts, memory, and tools. Its system-level view helps identify where updates happen and how to make them controllable.
Autodata generates training data with agent loops
Published on July 6, 2026, the Autodata method uses weak and strong solvers to select questions that distinguish model performance.
Automating scientific review while keeping human oversight
The article, published on June 29, 2026, frames scientific review at scale as a systems challenge and proposes four levels of collaboration between AI and humans.
BINEVAL makes LLM evaluations more interpretable
Published on June 27, 2026, the post describes a binary-question evaluation that matched or outperformed UniEval and G-Eval on three datasets without training.
- 19 days ago openreview.net
How agents can learn from interaction traces
A survey published on June 27, 2026, discusses how agents in production turn interaction traces into durable capabilities, from self-evolution to meta-evolution.
How to evaluate AI agent memory
An article published on June 26, 2026 examines agent memory systems and notes that existing evaluations often treat memory as a black box.
How to verify rewards for coding agents
A paper published on July 11, 2026 examines four reward-verification designs and finds no single solution to the challenge of reliably evaluating coding agents.
DiscoBench tests search agents on ambiguous requests
Published on June 29, 2026, DiscoBench evaluates whether language-model search agents can handle vague, underspecified, or factually incorrect requests in multi-step retrieval and reasoning tasks.
Dockerless proposes verifying patches without execution
Published on June 30, 2026, the paper describes a verifier that assesses coding-agent patches without running them. The proposal targets selecting SFT trajectories and providing rewards for RL, as an alternative to setting up repository-specific environments.
EvoRec proposes agents to optimize recommendations
Published on June 30, 2026, the paper presents a multi-agent framework intended to reduce manual optimization of recommendation systems by retaining methods across iterations and exploring ideas beyond a predefined search space.
Git as shared memory for research agents
A post published on September 17, 2026 describes agents recording verifiable contributions in a searchable Git history to share results and reduce repeated experiments.
HarnessBank: evolving agent harnesses
An article published on July 17, 2026 describes a method for improving agents by changing prompts, tools, and control loops around the model.
HASTE shows how to organize skills for ML agents
The material on HASTE reports that, across eight competitions, hierarchical skill loading achieved a 100% medal rate, compared with 62.5% for flat loading.
JAZ explores agents built around one code primitive
Published on September 26, 2026, the post describes the JAZ framework and reports results on StuLife and AppWorld, with lower costs than the compared approaches.
JIT-Agent adapts agent harnesses to each task
The article on JIT-Agent describes on-demand synthesis and evolution of harnesses for LLM-based agents, using modules for memory, planning, actions, and tools. Published August 27, 2026.
MemoHarness adapts agents using execution history
Published on July 17, 2026, the work presents a method that adjusts six dimensions of an agent harness using prior executions. On a shell-agent benchmark, it scored 0.806, compared with 0.722 for the strongest fixed-harness baseline.
Agent memory with a frozen model
In a post published on August 26, 2026, Recuris describes two memory loops and a fixed meta-agent that proposes repairs based on execution traces, without changing model weights.
Proactive memory for long-horizon agents
A paper describes a separate memory agent that maintains structured records and decides when to present reminders to an action agent. The approach was evaluated on Terminal-Bench 2.0 and tau-squared-Bench.
Recurrent memory for long-context agents
A paper published on September 29, 2026 identifies retention as a bottleneck in recurrent-memory agents and presents a training-free approach that selects one memory head and preserves the others.
Byte and token models: scaling results up to 1 trillion bytes
A paper published on September 14, 2026 compares distilled byte and token models and reports a higher performance ceiling for byte models as compute increases.
arxiv.orgLet's assess your project
I reply with scope and a proposal.