How Memory Shapes Autonomous LLM Agents
A survey of how memory is built, measured, and used in autonomous LLM agents, with a focus on design choices and open problems.
Timeline
The paper shows that LLMs often violate basic probabilistic consistency when aggregating subpopulation estimates.
E3 teaches LLM agents to estimate task scope first, then expand only when verification fails.
Direct-OPD lifts Qwen3-1.7B from 48.3% to 62.4% on AIME 2024 by distilling RL gains from a weaker model.
Build a layered evaluation pipeline for fine-tuned LLMs using task metrics, judges, safety checks, and human review.
DeepSpec is best understood as a conversation regeneration pipeline for training stronger models.
PAW compiles natural-language task specs into small local neural artifacts that run cheaply and offline.
LACUNA adds ground-truth parameter-level localization to test whether unlearning really removes memorized data.
A new benchmark shows AI coding agents can hide attacks across PRs and time their payloads to evade monitors.
This paper uses natural-language critiques to train policies from suboptimal demonstrations.
A layer-wise RL study finds that training one transformer layer can recover most post-training gains.
BINEVAL splits LLM evals into yes-or-no questions, improving inspectability and matching or beating G-Eval and UniEval on key benchmarks.
A new RL method uses metacognitive feedback to make LLMs express uncertainty more faithfully.
QVal is a training-free way to compare dense supervision signals for long-horizon LLM agents.
Fixed explanation datasets can still teach models to describe their current behavior, even as that behavior changes.
WorldEvolver updates an LLM agent’s test-time memory to improve foresight and planning without changing model weights.
LeVo 2 uses hierarchical modeling and progressive post-training to improve full-length song generation.
VLK synthesizes vision-language-kinematics supervision to train humanoid loco-manipulation from reconstructed indoor scenes.
I break down On-Policy Distillation and turn the idea into a copy-ready post-training template.
Google DeepMind’s science tools show how Google is packaging AI for researchers who want precision, not hype.
A new framework tests whether an LLM’s behavior transfers across payoff-equivalent decision environments.
Prompt injection lets hidden text steer LLMs, and recent tests show models like DeepSeek-R1 can be tricked at worrying rates.
Different zero-sum game solvers can converge to different Nash equilibria, and the choice is algorithm-dependent.
A new result characterizes when proper learning from positive-only samples is possible.
DexCompose composes pretrained hand policies into multi-task manipulation by assigning finger-level action ownership.
HaWoR’s hand reconstruction setup boils down to predicting MANO parameters, not raw meshes.
USC is advertising NVIDIA’s $30,000 academic grant for health and AI research, with June 30, 2026 applications due.
CUDA Toolkit 13.3 fixes a compiler bug from 12.8 that could corrupt registers in deeply divergent GPU kernels.
EAGLE3 is the main reason Kimi-K2.5-W4A8 decodes faster on AMD MI325X, not kernel tweaks.
A practical breakdown of enterprise LLM fine-tuning, from data prep to model choice, plus a copy-ready template.
Rust learners should clone freely at first, then optimize once they understand the problem.
Mistral OCR 4 adds boxes, block labels, and confidence scores to OCR, with API pricing from $4 per 1,000 pages.
ArBG replaces flow-based Boltzmann generators with autoregressive modeling for faster, more scalable equilibrium sampling.
RiVER shows LLMs can improve from score-based tasks without ground-truth answers by calibrating rewards from execution feedback.
DanceOPD trains flow-matching image models to combine text-to-image and editing skills without them fighting each other.
Microsoft Research opened a Spring 2026 CFP for AI that helps teams work better, with awards around $50K to $75K.
A Zhihu roundup highlights three 2026.06.24 AI papers on code generation, real-time music, and rare-disease diagnosis.
A June 24 arXiv roundup highlights agent memory, tool-use signals, and conversational search papers that push practical NLP forward.
Self-distillation can boost pass@1 while quietly reducing rollout diversity and hurting out-of-distribution robustness.
RevengeBench tests whether LLMs can reconstruct hidden game policies from behavior and improve with custom probes.
A two-stage training scheme gives VLA robots an explicit motion prior before cross-modal alignment.
I break down OPSD into a copyable loop for turning implicit user feedback into targeted correction and continual training.
UltraQuant shows 4-bit KV caching can speed long, multi-turn agent serving while keeping more context resident.
FLUX3D improves image-to-3D Gaussian generation by aligning sparse 3D latents with dense 2D image tokens.
InSight makes vision-language-action policies learn new manipulation skills without human demos of those target tasks.
Anthropic’s warning is justified, but the bigger problem is that AI control is already slipping beyond easy governance.
OpenAI researchers found multiple exploitable browser bugs in Chrome, Safari, and Firefox within a week.
OpenAI's LifeSciBench gives teams a better way to test life-science model quality.
CoorDex turns humanoid body and hand control into latent priors so dexterous manipulation can happen while the robot is moving.
Randomized YaRN helps LLMs generalize better from short training contexts to much longer reasoning windows.
AutoDex automates real-world dexterous grasp trials and labels physical outcomes without human intervention.