Papers.Cool Daily arXiv Digest | 2026-05-28
Automatically curated from 20 of the latest arXiv AI/ML papers, with 8 selected for in-depth translation and annotation.
---
ScientistOne: Towards Human-Level Autonomous Research via Chain-of-Evidence
arXiv: 2605.26340 | Authors: Rui Meng, Bhavana Dalvi Mishra, Jiefeng Chen et al. | Categories: cs.AI, cs.CL, cs.MA
Autonomous research agents have a fatal flaw: their papers look professional but are riddled with fabricated citations, unreproducible scores, and method descriptions disconnected from their code. This paper makes three contributions:
1. Chain-of-Evidence (CoE): a verifiability framework requiring every claim to be traceable to its evidence source 2. ScientistOne: an end-to-end autonomous research system that maintains the evidence chain across literature review, approach discovery, and manuscript writing 3. CoE Audit: four integrity checks — score verification, specification violation detection, citation verification, and method-code alignment
In an audit of 75 papers across 5 systems and 5 frontier research tasks, every baseline exhibited at least one systematic failure mode: fabricated citation rates up to 21%, score verification pass rates as low as 42%, and method-code alignment fluctuating between 20% and 80%.
ScientistOne achieved zero fabricated citations (0/337), perfect score verification (12/12), and the highest method-code alignment (14/15), while matching or exceeding human expert performance on all five tasks.
---
Your Agents Are Aging Too: Agent Lifespan Engineering for Deployed Systems
arXiv: 2605.26302 | Authors: Jianing Zhu, Yeonju Ro, John Robertson et al. | Categories: cs.AI, cs.CL, cs.MA
Long-lived AI agents are deployed as persistent operational systems, yet are evaluated like freshly initialized models. Day-one benchmarks ignore a basic question: how long does an agent remain reliable after deployment?
Even with frozen model weights, an agent's effective state keeps changing — compressing interaction history, retrieving from growing memory stores, revising facts, undergoing routine maintenance. Reliability is therefore a lifespan-level property, not a snapshot property of the base model.
The paper introduces AgingBench, a longitudinal reliability benchmark, identifying four aging mechanisms:
- Compression aging: information loss from history compression
- Interference aging: new and old memories interfering with each other
- Revision aging: inconsistencies introduced by fact updates
- Maintenance aging: side effects of system maintenance
- PolyFusion: aligns complementary views of millions of polymers — sequences, topology, 3D geometry, and fingerprints — learning a transferable shared latent space that improves thermophysical property prediction and enables chemically valid, structurally novel generation beyond reference design spaces
- PolyAgent: links prediction and inverse design to evidence retrieval over the polymer literature, proposing, evaluating, and contextualizing hypotheses with explicit precedents in one workflow
Across 7 scenarios, 14 models, multiple memory strategies, and ~400 runs (8–200 sessions), behavioral tests may look fine while factual accuracy has already decayed — and identical wrong answers may require entirely different repair strategies.
---
Can LLMs Introspect? A Reality Check
arXiv: 2605.26242 | Authors: Shashwat Singh, Tal Linzen, Shauli Ravfogel | Categories: cs.AI
Can LLMs detect and report their own internal states? Several studies say yes. Drawing on lessons from human metacognition research, this paper argues that conclusion may be premature: genuine introspection must be distinguished from pattern matching on surface-level cues, and behavioral evidence alone is insufficient to establish strong introspective claims.
The authors re-examine two evaluation paradigms: 1. State tampering detection: models cannot reliably distinguish interventions on internal states from manipulations of the input, suggesting success reflects anomaly detection rather than awareness of internal states 2. Hidden-state label prediction: classifiers with input-only access match the model's own in-context predictions, suggesting no privileged access to internal representations
In relabeled controlled settings where models cannot rely on task semantics, performance approaches random chance.
---
The MiniMax-M2 Series: Mini Activations Unleashing Max Real-World Intelligence
arXiv: 2605.26494 | Authors: MiniMax, Aili Chen, Aonian Li et al. | Categories: cs.AI, cs.CL, cs.LG
MiniMax introduces the M2 series of Mixture-of-Experts language models under the philosophy: mini activations unleash max real-world intelligence. The flagship M2 has 229.9B total parameters with only 9.8B activated per token, designed end-to-end for agentic deployment on three pillars:
1. Agent-driven data pipelines: generating large-scale, verifiable agentic coding and agentic cowork trajectories grounded in executable workspaces and artifact-aligned rewards 2. Forge: a scalable agent-native RL system for long-horizon agent trajectories, featuring window-FIFO scheduling, prefix-tree merging, inference optimization, and a training-inference-agent decoupled architecture supporting white-box and black-box agents 3. M2.7: an early step toward self-evolution — autonomously debugging training runs and modifying its own scaffold
From M2 to M2.7, the series reaches frontier performance on agentic coding, deep search, office tasks, and reasoning benchmarks.
---
JobBench: Aligning Agent Work With Human Will
arXiv: 2605.26329 | Authors: Yuetai Li, Yichen Feng, Zhangchen Xu et al. | Categories: cs.AI
Current occupational AI agent benchmarks are scoped by economic value, telling a replacement story. JobBench shifts the view to evaluating agents on tasks human experts identify as high-priority for delegation — empowering humans rather than replacing them.
It covers 130 agentic tasks across 35 occupations. Each task is packaged as a workspace of heterogeneous reference files, requiring reasoning through the cluttered information flow of real professional work. Outputs are scored by fact-grounded rubrics averaging 35.6 binary criteria per task.
Thirty-six models were evaluated; the strongest, Claude Opus 4.7 under Claude Code, reached only 45.9%. JobBench aims to shift community goals from replacement to augmentation: building agents that do what humans actually want to delegate, not just what is most economically valuable.
---
Is Agent Memory a Database? Rethinking Data Foundations for Long-Term AI Agent Memory
arXiv: 2605.26252 | Authors: Abdelghny Orogat, Essam Mansour | Categories: cs.AI, cs.DB
Long-running agents need persistent memory for cross-session learning, reduced context re-injection, and decision auditing. Current agent memory systems and database paradigms treat memory as storage, localizing correctness at records, embeddings, or edges. This yields four recurring failure modes: unregulated growth, missing semantic revision, capacity-driven forgetting, and read-only retrieval without write-integrity guarantees.
The paper proposes Governed Evolving Memory (GEM): correctness as a property of state trajectories rather than individual records, replacing record-level operations with four state-level operators — ingest, revise, forget, retrieve — constrained by six correctness conditions. Feasibility is validated in the MemState prototype, built on a property-graph backend.
---
OmniToM: Benchmarking Theory of Mind in LLMs via Explicit Belief Modeling
arXiv: 2605.26322 | Authors: Adam Bawatneh, Sagar Sapkota, Amrit Singh Bedi et al. | Categories: cs.AI
Theory of Mind (ToM) in LLMs is usually evaluated via end-point question answering, which obscures whether the model actually constructs the mental-state representations needed for robust social reasoning.
OmniToM evaluates these representations directly, requiring explicit belief structures for all relevant actors in a narrative, where each belief is a minimal proposition about what an actor takes to be true of the world or of another actor's mental state. Evaluation proceeds in two stages: belief extraction from stories, then belief labeling along a seven-dimensional schema (recursion order, truth status, knowledge acquisition, explicitness, content type, mental source, context).
Built from 895 stories and 22,343 labeled belief propositions, zero-shot evaluation reveals an actor-specific belief tracking bottleneck: current LLMs struggle with the knowledge-acquisition and representation decisions needed to convert narrative facts into actor beliefs and shared mental states.
---
PolyFusionAgent: A Multimodal Foundation Model and Autonomous AI Assistant for Polymer Property Prediction and Inverse Design
arXiv: 2605.26543 | Authors: Manpreet Kaur, Xingying Zhang, Qian Liu | Categories: cs.AI, cs.LG
Polymer discovery is central to fields from energy storage to biomedicine, but is hindered by an astronomically large chemical design space and fragmented structure/property/prior-knowledge representations, leaving many AI models disconnected from physical and experimental reality.
PolyFusionAgent couples a multimodal polymer foundation model (PolyFusion) with a tool-augmented, literature-grounded design agent (PolyAgent):
---
*Automatically curated on 2026-05-28*