Veritas harness auto-generates agent-task verifiers that beat seven alternatives
With the model fixed, generated verifiers matched official verdicts better than seven harnesses on all four execution-scored benchmarks.
Stories per day · 30 days
3 RSS
Relatedbenchmarks5reinforcement-learning4embodied-ai3llm-agents3agent-memory1agents1aime1autoregressive1auxiliary-rewards1coding-agents1diffusion1error-reproduction1game-ai1generalist-ai1hearthstone1hypothesis-generation1llm1llm-compression1
49 stories, newest first · one card per story
Stories tagged #self-improvement
Generalist robot policies improved via reinforcement learning in multi-task settings
Generalist policies enable online multi-task improvement with generalization to unseen tasks.
arXiv study shows coding agents can compress LLMs
29 of 36 agent submissions pass verification, achieving 77.33% downstream accuracy on compressed models.
PyRUA-REAL enables robot exploration and adaptive learning through recursive self-improvement
The PyRUA-REAL framework allows coding agents to retain interaction experience as persistent memory, achieving 91.1% success on eight real-world tasks.
Auxiliary Rewards Improve Small-Language-Model Self-Improvement
UID-Conj. auxiliary rewards boost online adaptation and offline SFT by 2.08 and 1.93 points over baselines on AIME math tasks.
A-RSI refreshes tasks for zero-data multimodal self-improvement
On an eleven-benchmark suite the authors report the average rising from 57.77 to 61.05 using 800 fresh tasks.
The Trinity evolves multimodal-agent harnesses from shared capabilities
The authors report mean gains of 34 points on 28 training benchmarks and 26 points on 18 unseen benchmarks.
AutoTavern agents build a Hearthstone AI that rivals top ladder players
In 20 days the system produced TavernGym and 23 TavernAgent versions whose last 20 games averaged 1.60 placement.
Stored abstractions can reproduce errors in LLM agent memory
Anti-Persistence rebuilds task guidance on demand and discards it, with the lowest average LLM-call cost tested.
Paper studies physical recursive self-improvement in VLM agents
PhysRSI-eval scores 18 manipulation tasks and finds close public averages can hide late or inefficient gains.
DiscoveryRSI recursively improves scientific hypothesis generation
Automated scores rose from 4.86 to 5.39 after two self-improvement rounds, with larger gains in falsifiability and data resolvability.
Language Models Learn to Improve Language Models
LM2LM uses a research world model as an RL reward so Qwen3.8-27B proposes stronger model improvements than GPT-6-sol.
Ornith-1.5-122B mixture-of-experts model trends on Hugging Face
The card describes a 122B MoE for single-GPU use and a quantized mobile variant, with vendor benchmark scores.
1
RSIFlow adapts agent self-improvement with an evolving meta-agent
Authors report state-of-the-art average scores across six benchmarks and two backbone models.
HACE coevolves specialist models and LLM agents
The method beat evaluated self-improving baselines on recommendation, protein function, and crystal stability tasks.
SIMIT Self-Improves Vision-Language Models
SIMIT generates query-specific practice examples at test-time yielding +7.20% mean gain over BAGEL-7B.
MC-EvoBench: Auditing Experience-Driven Self-Improvement in Minecraft Agents
MC-EvoBench tests whether Minecraft agents convert experience into durable capabilities that transfer across reset worlds.
Google Research RRSI Guide: Mastering Self-Improving AI Agents
Google Research's RRSI uses noise bands, cost rules, and leakage screens for safer self-improving AI agents.
RoboRSI: Stable, efficient, and reusable robot self-evolution in complex real-world environments
RoboRSI lets a mobile manipulator reuse validated skills, beating baselines by up to 11 points.
VeriFine: Scaling Verification for Self-Improvement in Embodied Reasoning
VeriFine co-evolves an embodied policy, curriculum, and judge so verification keeps pace with new failure patterns.
Agent Plasticity: Measuring Self-Improvement Through Experience
Researchers define agent plasticity to measure how efficiently AI agents improve from experience.
Co-Evolving Robot Orchestrators and Policies through Deployment
Robo-COP co-evolves robot policies and orchestrators in deployment, raising real-world success from 38.3% to 50%.
VeriFine: Scaling Verification for Self-Improvement in Embodied Reasoning
VeriFine co-evolves policies, curricula, and judges so embodied agents keep self-improving as failure patterns change.
EmbodiedSmith: Scaling Embodied Data through Recursive Self-Improvement Flywheel in Simulation
EmbodiedSmith generates diverse robot simulation data by recursively co-refining scenes and tasks for foundation-model training.