Public story · 2026-02-28 · source-backed

EMPO2: Memory-Augmented Agent (ICLR 2026)

Issue

2026-02-28

Confidence

source-backed

Sources

Redaction

redacted

Story

Hybrid on/off-policy RL giving agents non-parametric memory for exploration. 128.6% improvement over GRPO on ScienceWorld, 11.3% on WebShop. Agents generalize to out-of-distribution tasks with "only a few trials with memory and no parameter updates." (arXiv)

SkillPyramid organizes agent skills into a hierarchy
SkillPyramid extends Voyager-style libraries with a hierarchical topology plus a self-evolution loop, raising average reward 38% and cutting execution steps 27.7% across ALFWorld, WebShop, and ScienceWorld. The takeaway: don't append every successful trajectory as a new flat s...
Memory-R1: RL-Based Agent Memory Management
Two specialized RL agents — Memory Manager (ADD/UPDATE/DELETE operations) and Answer Agent — fine-tuned with PPO and GRPO. With only 152 training QA pairs, outperforms baselines across three benchmarks. Directly applicable to persistent agent memory systems. arXiv 2508.19828
Unsloth makes single-H100 long-context RL fine-tuning actually tractable.
New batching algorithms enable ~7x, up to 12x+, longer-context GRPO training with no accuracy or speed penalty versus optimized FA3 and chunked-loss setups (Unsloth Docs). Qwen3-8B GRPO reaches 110K context on one 80GB H100 via vLLM plus QLoRA. For solo builders doing reasonin...
HCAPO: Hindsight Credit Assignment for Sparse-Reward Agent RL.
LLM as post-hoc critic for step-level Q-values. +7.7% WebShop, +13.8% ALFWorld over GRPO. Third paper in the online RL-for-agents cluster this week. arXiv:2603.08754
Gemma 4 Drops Four Open Models, and the 31B Dense Beats a 397B MoE
Google DeepMind released Gemma 4 on April 2 with four model sizes (E2B, E4B, 26B MoE, 31B Dense) under Apache 2.0. Multimodal (text, vision, audio). 256K context. Native thinking and tool-calling optimized for agentic workflows. Day-zero ecosystem support across vLLM, llama.cp...
CogEvol reports a caught reward-hacking episode alongside its production numbers.
It turns a course brief into finished slides or a self-contained interactive HTML page in one pass (arXiv 2608.30968). Across 220k production requests the median slide takes 17 seconds and an interactive page 59. A hybrid rule-plus-VLM reward drives GRPO, hardened after the te...

Source trail

arxiv.org/abs/2602.23008 More from arxiv.org

Entities

Augmented Agent EMPO2 GRPO Hybrid ICLR Memory ScienceWorld WebShop

Provenance

Canonical issue: Ramsay Research Agent — 2026-02-28
AI generated: no
Story unit: 2026-02-28-empo2-memory-augmented-agent-iclr-2026
Labels: source-backed, canonical briefing excerpt

Fetching from the wire…

EMPO2: Memory-Augmented Agent (ICLR 2026)

Story

Related stories

SkillPyramid organizes agent skills into a hierarchy

Memory-R1: RL-Based Agent Memory Management

Unsloth makes single-H100 long-context RL fine-tuning actually tractable.

HCAPO: Hindsight Credit Assignment for Sparse-Reward Agent RL.

Gemma 4 Drops Four Open Models, and the 31B Dense Beats a 397B MoE

CogEvol reports a caught reward-hacking episode alongside its production numbers.

Source trail

Entities

Provenance

EMPO2: Memory-Augmented Agent (ICLR 2026)

Story

Related stories

SkillPyramid organizes agent skills into a hierarchy

Memory-R1: RL-Based Agent Memory Management

Unsloth makes single-H100 long-context RL fine-tuning actually tractable.

HCAPO: Hindsight Credit Assignment for Sparse-Reward Agent RL.

Gemma 4 Drops Four Open Models, and the 31B Dense Beats a 397B MoE

CogEvol reports a caught reward-hacking episode alongside its production numbers.

Source trail

Entities

Provenance

Story

Related stories

SkillPyramid organizes agent skills into a hierarchy

Memory-R1: RL-Based Agent Memory Management

Unsloth makes single-H100 long-context RL fine-tuning actually tractable.

HCAPO: Hindsight Credit Assignment for Sparse-Reward Agent RL.

Gemma 4 Drops Four Open Models, and the 31B Dense Beats a 397B MoE

CogEvol reports a caught reward-hacking episode alongside its production numbers.

Tomorrow's issue, in your inbox.

Source trail

Entities

Provenance

Story

Related stories

SkillPyramid organizes agent skills into a hierarchy

Memory-R1: RL-Based Agent Memory Management

Unsloth makes single-H100 long-context RL fine-tuning actually tractable.

HCAPO: Hindsight Credit Assignment for Sparse-Reward Agent RL.

Gemma 4 Drops Four Open Models, and the 31B Dense Beats a 397B MoE

CogEvol reports a caught reward-hacking episode alongside its production numbers.

Tomorrow's issue, in your inbox.

Source trail

Entities

Provenance