arXiv AI Papers Digest — 2026-05-28
Eight selected AI/ML papers from arXiv, covering autonomous research, agent reliability, LLM introspection, efficient MoE models, agent benchmarks, memory systems, Theory of Mind, and scientific AI.
---
1. ScientistOne: Towards Human-Level Autonomous Research via Chain-of-Evidence
- arXiv: 2605.26340 | cs.AI, cs.CL, cs.MA
- arXiv: 2605.26302 | cs.AI, cs.CL, cs.MA
- Compression aging: information loss from history compression
- Interference aging: new and old memories interfering
- Revision aging: fact updates introducing inconsistencies
- Maintenance aging: side effects of routine maintenance
- arXiv: 2605.26242 | cs.AI
- arXiv: 2605.26494 | cs.AI, cs.CL, cs.LG
- arXiv: 2605.26329 | cs.AI
- arXiv: 2605.26252 | cs.AI, cs.DB
- arXiv: 2605.26322 | cs.AI
- arXiv: 2605.26543 | cs.AI, cs.LG
- PolyFusion: aligns sequences, topology, 3D geometry, and fingerprints of millions of polymers in a transferable shared latent space, improving thermophysical property prediction and enabling chemically valid, structurally novel generation
- PolyAgent: links prediction and inverse design through evidence retrieval from polymer literature, proposing and contextualizing hypotheses with explicit precedents
Autonomous research agents produce professional-looking manuscripts riddled with verifiability failures: fabricated citations, unreproducible scores, and method descriptions diverging from implementations. 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 maintaining evidence chains across literature review, solution discovery, and manuscript writing 3. CoE Audit: four integrity checks — score verification, specification violation detection, citation verification, method-code alignment
Auditing 75 papers across 5 systems and 5 frontier tasks, all baselines exhibited at least one systematic failure mode: fabricated citation rates up to 21%, score verification pass rates as low as 42%, method-code alignment ranging from 20%–80%. ScientistOne achieved zero fabricated citations (0/337), perfect score verification (12/12), and the highest method-code alignment (14/15), matching or exceeding human expert performance on all five tasks.
---
2. Your Agents Are Aging Too: Agent Lifespan Engineering for Deployed Systems
Long-lived agents are deployed as persistent operational systems but evaluated like freshly initialized models. Even with frozen weights, an agent's effective state keeps changing. The paper introduces AgingBench, a longitudinal reliability benchmark identifying four aging mechanisms:
Across 7 scenarios, 14 models, multiple memory strategies, and ~400 runs (8–200 sessions): behavioral tests may pass while factual accuracy has decayed, and identical wrong answers may require entirely different remediation strategies.
---
3. Can LLMs Introspect? A Reality Check
Drawing on human metacognition research, the authors argue that claims of LLM introspection may be premature: behavioral evidence alone cannot distinguish genuine introspection from pattern matching on surface cues. Re-examining two paradigms:
1. State-tamper detection: models cannot reliably distinguish interventions on internal states from input manipulations, suggesting anomaly detection rather than genuine awareness of internal states 2. Hidden-state label prediction: classifiers with input-only access match the model's in-context predictions, indicating no privileged access to internal representations
In relabeled control settings where models must rely on internal representations rather than task semantics, performance approaches chance.
---
4. The MiniMax-M2 Series: Mini Activations Unleashing Max Real-World Intelligence
MiniMax-M2 is a family of Mixture-of-Experts models built on the principle that mini activations can unleash maximum 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:
1. Agent-driven data pipelines: large-scale verifiable trajectories for agentic coding and agentic cowork, grounded in executable workspaces and artifact-aligned rewards 2. Forge: a scalable agent-native RL system for long-horizon trajectories, with window-FIFO scheduling, prefix-tree merging, inference optimization, and train-inference-agent decoupling 3. M2.7: early steps toward self-evolution — autonomously debugging training runs and modifying its own scaffold
The series achieves frontier performance on agentic coding, deep search, office tasks, and reasoning benchmarks.
---
5. JobBench: Aligning Agent Work With Human Will
Most occupational agent benchmarks tell a replacement story driven by economic value. JobBench instead evaluates agents on workflows experts identify as high-priority for delegation — empowerment, not replacement. It covers 130 agentic tasks across 35 occupations, each packaged as a workspace of heterogeneous reference files, scored via fact-grounded rubrics averaging 35.6 binary criteria per task. Among 36 evaluated models, the strongest (Claude Opus 4.7 under Claude Code) reaches only 45.9%.
---
6. Is Agent Memory a Database? Rethinking Data Foundations for Long-Term AI Agent Memory
Current agent memory systems treat memory as storage, localizing correctness at record, embedding, or edge level. This yields four recurring failure modes: unregulated growth, missing semantic revision, capacity-driven forgetting, and read-only retrieval. The paper proposes Governed Evolving Memory (GEM), treating correctness as a property of state trajectories rather than individual records, with four state-level operators (ingest, revise, forget, retrieve) constrained by six correctness conditions. Validated in the MemState prototype on a property-graph backend.
---
7. OmniToM: Benchmarking Theory of Mind in LLMs via Explicit Belief Modeling
Theory of Mind in LLMs is usually assessed via end-point QA, obscuring whether models actually build mental-state representations. OmniToM requires explicit belief structures for all relevant actors, evaluated in two stages: belief extraction and belief labeling across seven dimensions (recursion order, truth status, knowledge access, 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: LLMs struggle to convert narrative facts into actor beliefs and shared mental states.
---
8. PolyFusionAgent: A Multimodal Foundation Model and Autonomous AI Assistant for Polymer Property Prediction and Inverse Design
PolyFusionAgent couples a multimodal polymer foundation model with a tool-augmented, literature-grounded design agent:
*Auto-collected 2026-05-28*