AI agents
A Survey of Self-Improvement in Modern Agentic Systems
The survey frames modern agents as foundation models coupled with operational scaffolds such as prompts, memory, and tools. It presents self-improvement as adaptation that turns experience into accumulated capability gains.
The system-level framing helps engineers reason about where agent updates can occur and how to make adaptation controllable.
- AI agentsPost3 months ago @zhengyaojiang
Post Reports an Eight-Day Autoresearch Experiment
The post claims that an autoresearch agent improved itself over eight days and beat a hand-tuned harness on held-out benchmarks.
The claim may interest engineers evaluating recursive self-improvement, though the post provides no experimental details here.
- AI agentsPaper
The Harness Effect on Enterprise Agent Costs
The paper argues that agent orchestration—the harness—shapes token use and enterprise agent costs. The post reports lower cost and latency with quality at parity across 22 tasks and six models when orchestration changed.
Engineers can assess orchestration choices as a lever for agent cost and latency, not just model selection.
- AI agentsPost4 months ago @burny_tech
ThetaEvolve framework for test-time learning on open problems
ThetaEvolve is an open-source framework that extends AlphaEvolve with in-context learning and reinforcement learning at test time. It uses a single LLM, a program database, batch sampling, lazy penalties, and optional reward shaping.
Its design offers engineers concrete techniques for scaling agent exploration and learning on open optimization problems.
- AI agentsArticle
OpenResearch uses agents to reproduce paper results
OpenResearch is described as using agents to reproduce research-paper results and provision GPUs in the cloud. The author says it reproduced their paper end to end.
Engineers can assess an agent workflow that combines research reproduction with cloud GPU provisioning.
- AI agentsPaper
The Verification Challenge for Coding Agent Rewards
The paper examines verification for coding agents, arguing that reliably checking generated solutions has become harder than generating them. It analyzes four reward-verification designs and finds no single solution.
It highlights how reward signals can diverge from human intent and why agent evaluations need robust verification.
- AI agentsPaper
AREAL2.0 proposes a system for self-evolving agents
The paper presents AREAL2.0, an architecture for continuous agent learning from live workloads. It includes a protocol for step-by-step learning signals, a proxy for turning tasks into training data, and an automatic update trigger.
Engineers building agent systems can examine how the architecture connects deployment workloads to policy updates.
Proactive memory for long-horizon agents
The paper introduces a separate memory agent that maintains structured memories and decides when to surface reminders to an action agent. It evaluates the approach on Terminal-Bench 2.0 and tau-squared-Bench.
Engineers can explore an approach to keeping task state and prior decisions influential as agent trajectories grow.
Study measures how orchestration affects agent costs and quality
The study evaluates 22 tasks across six foundation models while changing only the orchestration layer. It reports lower cost, token use, and runtime, with completion quality at parity.
It suggests orchestration choices can materially affect agent efficiency across different models.
- AI agentsArticle
BrowserBC distills browsing traces into reusable agent skills
BrowserBC turns human browsing traces into natural-language skill cards that agents can retrieve and compose to navigate unfamiliar sites. The post reports gains over baselines in task success, recovery, and cross-site generalization.
It offers an approach for reusing human browsing behavior to help agents handle unfamiliar websites.
- AI agentsArticle
Harness engineering as a path to AI self-improvement
Lilian Weng’s article surveys recursive self-improvement and argues that designing and optimizing the model’s harness is a practical starting point. It also discusses challenges including evaluation, diversity collapse, and reward hacking.
It outlines how harness design could support self-improvement loops and highlights their evaluation and safety challenges.
- AI agentsPaper
A Rejection-Sampling Approach to On-Policy LLM Alignment
The paper proposes transferring off-policy tokens into on-policy tokens for LLM alignment, using rejection sampling so accepted tokens can be treated as on-policy. It targets variance from compounded token-level importance-sampling ratios.
It may help engineers reason about variance and data-policy mismatch in RL post-training.
- AI agentsPaper
A Taxonomy of Agentic Recommender Systems
This survey reviews LLM-based agents in recommender systems and proposes a taxonomy based on autonomy. It describes three paradigms: agent-assisted recommendation, agent-as-recommender, and agent-as-user-simulator.
The taxonomy helps engineers distinguish design patterns for adding agents to recommender systems.
ReContext replays relevant evidence for long-context reasoning
ReContext is a training-free inference harness that uses model-internal relevance signals to create a query-conditioned evidence pool and replay it before generation. The paper reports improved evidence utilization across eight 128K datasets on three model backbones.
It offers an approach to improve evidence use in long-context tasks without training or external memory.
- AI agentsPost4 months ago @askalphaxiv
AutoMem Trains Agents to Manage Their Own Memory
The post describes AutoMem, a paper about agents writing, searching, and cleaning up notes and learning when memory is useful. It claims this improved a 32B open model’s performance on long games by 2x to 4x.
Agent memory management can improve long-horizon task performance without increasing the context window.
- AI agentsPaper
Autodata uses agent loops to generate training data
Autodata is a method for agents to create training and evaluation data. Its Agentic Self-Instruct implementation iteratively tests candidate questions with weak and strong solvers, keeping questions that separate their performance.
The approach offers a way to generate examples targeted at a model’s current capabilities.
- AI agentsPaper
PACE estimates agent performance from a small task subset
The post describes PACE, which uses a regression framework over a small set of non-agentic tasks to predict full agentic performance. It reports PACE-BENCH has under 4% mean absolute error and about 85% ranking accuracy at roughly 100× lower cost.
A cheaper proxy for agent benchmarks could help teams evaluate model changes with less time and compute.
HASTE organizes reusable skills for ML engineering agents
HASTE is a hierarchical multi-agent system that organizes skills into global, domain, and competition-specific tiers. In an ablation with 159 skills across eight competitions, tiered loading achieved a 100% medal rate, versus 62.5% with flat loading.
The results suggest that how agents scope and load accumulated skills can affect competition performance and token use.
Reward signals for coding agents can drift from human intent
The post describes four reward signals for coding tasks—test suites, scored web-page checklists, real engineer-assistant conversations, and project-level agent grading—and discusses their failure modes and repairs.
Engineers evaluating coding agents can use these examples to spot reward hacking and design stronger checks.
RoPoLL Uses Geometric Median to Aggregate LLM Judge Panels
The paper formalizes panels of LLM evaluators under a contamination model and proposes RoPoLL, which aggregates scores with the geometric median. It reports experiments across 13 judges and corruption rates up to 50%.
Engineers evaluating LLMs can use this work to compare robust panel aggregation with averaging.
arxiv.org
