# E91 Research Review

Review date: 2026-11-18, by session `ec`.

## What was actually read

- **Fetched, abstract or landing page (2026-11-18):** CL-Bench, MemoryArena, AMA-Bench, AgentWorld, the long-horizon architecture paper, and AgentLAB.
- **Search-result summaries only:** CLE, the generalization papers, the science-agent benchmarks, and the agentic-RL papers. Their numbers are quoted as the search summary stated them and are marked **[summary]**. None were verified against full text.
- **No full paper text was read.** Where methodology or limitations are not stated on the abstract page, this review says so instead of filling them in.

## How the review is organized

- Section 1 covers the external sources.
- Section 2 covers earlier CAIN evolutions, used as baselines.
- Section 3 lists what E91 takes from the review.

Status labels used below:

| Label | Meaning |
|---|---|
| PR | peer-reviewed or accepted |
| PP | preprint |
| BM | benchmark |
| SV | survey |

## 1. External sources

### CL-Bench (PP, BM; ICML 2026 listing and accepted to NeurIPS 2026 per Snorkel)

- **Title:** Continual Learning Bench: Evaluating Frontier AI Systems in Real-World Stateful Environments
- **Authors:** Asawa, Glaze, Orlanski, Ramakrishnan, Xu, Biswal, Chen, Sala, Zaharia, Gonzalez
- **Date:** arXiv submission 2026-06-04
- **URL:** https://arxiv.org/abs/2606.05661
- **Question:** do LLM systems improve with experience in stateful environments?
- **Method:** six domains, each with a shared latent structure. A gain metric isolates learning from prior capability. Agent architectures are compared, from naive in-context learning (ICL) to dedicated memory systems.
- **Findings (abstract):** agents overfit to immediate observations or fail to reuse knowledge across instances. *"Naive ICL outperforms systems dedicated to memory management."*
- **Limitations:** not stated in the abstract.
- **Relevance:** E89 ran the official harness. Their Opus hybrid scored 0.478 (3 runs), versus ICL at 0.637 (1 run): 75% of ICL at about 23% of the cost.
- **Informs E91:** pre-registered H2 (memory-as-context does not help) and a G-vs-B/C comparison where B is the ICL-style arm.

### Continual Learning Exam (CLE) (PP, BM) [summary]

- **Date:** arXiv 2610.33956, late October 2026
- **URL:** https://papers.cool/arxiv/2610.33956
- **Design:** 120 human-authored tasks, each a five-level ladder.
- **Findings:** the best system clears 47.7% of levels and finishes all five levels in 23.3% of tasks. Feedback plus one revision gives a median gain of +5.3 pp across 17 configurations.
- **Not read:** authors, full methodology and limitations.
- **Informs E91:** the "verify, feed back, retry" strategy (`verify_retry`) and the expectation that feedback-driven gains are small.

### MemoryArena (PP, BM)

- **Authors:** He, Wang, Zhi, Hu, Chen, Yin, Chen, Wu, Ouyang, Wang, Pei, McAuley, Choi, Pentland
- **Date:** arXiv:2602.16313 (2026), per the page's BibTeX
- **URL:** https://memoryarena.github.io/
- **Question:** memory in interdependent multi-session agent tasks.
- **Method:** four task families (shopping, travel planning, progressive search, formal reasoning).
- **Findings (page):** agents that near-saturate LoCoMo perform poorly here.
- **Numbers:** none on the page.
- **Informs E91:** a decision is scored on later held-out tasks, not on recall of the stored item.

### AMA-Bench (PR, ICML 2026, PMLR 306)

- **Authors:** Zhao, Yuan, Huang, Yuan, Yu, Xu, Hu, Shankarampeta, Huang, Ni, Tian, Zhao
- **URL:** https://proceedings.mlr.press/v306/zhao26bs.html
- **Question:** long-horizon agent memory over real and synthetic trajectories.
- **Findings (abstract):** memory systems underperform because they fail to capture causal and objective information and rely on lossy similarity retrieval. Their AMA-Agent (causality graph plus tool-augmented retrieval) reaches 57.22%, "outperforming the strongest baseline by 11.16%". Whether that is pp or relative is not stated.
- **Informs E91:** G keys memory on *task structure and verified conditions*, not on text similarity. Arm C is the similarity baseline.

### AgentWorld (PP; accepted at COLM 2026 per arXiv comments)

- **Authors:** Shu, Zhang, Cho, Yang, Yuan, Zheng, Guntuku, Ungar, Yu, Zhang
- **Date:** 2026-09-25
- **URL:** https://arxiv.org/abs/2609.31590
- **Method:** 100 MMORPG tasks for 3–20 agents, with a Causal Collaboration Effectiveness metric.
- **Findings:** the best model reaches 52.0% task success. Failure modes are communication breakdowns, role confusion and lost shared plans.
- **Informs E91:** the team experiment (session 1b) measures duplicate work and false consensus, not just success.

### An Architecture for Long-Horizon Agents (PP, position and architecture)

- **Authors:** Nijkamp, Koul, Pakhomov, Pang
- **Date:** 2026-09-17
- **URL:** https://arxiv.org/abs/2609.19519
- **Claim:** an agent must run continually without forgetting before it can learn continually, and that capability lives in the harness. The architecture uses levels, clocked ticks, and cascaded escalation.
- **Evidence:** one ten-day case study, not a controlled comparison.
- **Informs E91:** persistence across process boundaries (H7, and session 1b's mission runtime). A single case is not treated as evidence of a general effect.

### AgentLAB (PR, ICML 2026, PMLR 306)

- **Authors:** Jiang, Wang, Liang, Wang
- **URL:** https://proceedings.mlr.press/v306/jiang26as.html
- **Design:** five long-horizon attack types (intent hijacking, tool chaining, task injection, objective drifting, memory poisoning), 28 environments, 644 test cases. Agents remain highly susceptible.
- **Informs E91:** H8 (forged experience rejected on read) and session 1b's security lab.

### Generalization and RL for agents [summary]

- **"Can RL Improve Generalization of LLM Agents?"** (arXiv 2603.12011): large in-domain gains, much smaller held-out gains. One AlfWorld figure is +78.62 in-domain versus +4.91 held-out.
- **"Paying Less Generalization Tax"** (arXiv 2601.18217): held-out performance can drop while in-domain scores rise.
- **MemCo** (arXiv 2610.07376): ALFWorld seen versus unseen splits.
- **Informs E91:** held-out results are reported separately from dev, and transfer has negative controls.

### Autonomous science agents [summary]

- **PRBench** (arXiv 2603.27646): physics paper reproduction. Failures include fabricated output data.
- **AutoExperiment** (ICLR 2026): progressive code masking.
- **AstaBench** (ICLR 2026): 2,400+ problems and 57 agents; the authors conclude AI is far from solving science research assistance.
- **AIRS-Bench:** 20 ML tasks.
- **MLReplicate:** every system it tested produced hallucinated results.
- **ResearchClawBench:** the best agent averaged 21.5.
- **Informs E91:**
  - `science.claim_check` makes fabricated citations a scored failure.
  - `science.hypothesis_test` makes REJECTED the correct answer in 2/3 of tasks.
  - `science.underpowered` makes INSUFFICIENT_EVIDENCE the correct answer.

### Agentic RL and self-improvement [summary]

- **AReaL2.0 / self-evolving agents** (arXiv 2607.01120): systems-level argument for online RL.
- **RetroAgent** (arXiv 2603.08561): hindsight reflection RL beats GRPO.
- **Agentic Self-Learning** (arXiv 2510.14253).
- **Survey:** arXiv 2604.27859.
- **Not implemented:** no weight updates happen in E91. "Learning" here means evidence-gated procedure selection, and the reports say so.

## 2. Earlier CAIN evolutions as baselines

| Evolution | State on 2026-11-18 | What E91 takes |
|---|---|---|
| E88 (`f2d12e8b6`) | committed. 94/94 tests and the learning experiment reproduced byte-identically by `ec`. | latent-rule consolidation helps, but injecting a procedure into a weak model hurt (C 4/15 vs H 0/15 in phase 2). The governance primitives are reused. |
| E89 (`081723055`, session 1b) | committed. CL-Bench official harness. | full-context ICL is the bar; memory helps mainly on cost. |
| E90 (uncommitted, owner session unreachable) | **UNVERIFIED** | two seeds: full arm 0.612/0.596 vs basic retrieval 0.579/0.577. The causal-only arm does not beat retrieval. The gain is mostly from abstention. |

## 3. What E91 takes from this

1. Measure learning on held-out tasks against an ICL-style baseline (B), not against no-memory.
2. Prefer structure- and condition-keyed memory over similarity (AMA-Bench).
3. Every transfer claim needs a negative control (generalization-tax findings).
4. Science tasks must reward rejection and abstention, and penalise fabricated sources.
5. Say plainly that no weights are trained and that procedures are designer-supplied.
