arXiv AI/ML Daily Digest - Sep 4, 2026: 20 New Papers

📚 arXiv AI/ML Paper Digest · 2026-09-04

> Automatically collected from arXiv cs.AI / cs.LG / cs.CL / cs.CV — 20 new papers.

---

🔥 Highlights

1. EvalDetectBench: Measuring LLMs' Evaluation Awareness

Authors: Benjamin Feuer, Lucas Rosenblatt, Oussama Elachqar arXiv: 2609.01775

As AI models evolve from simple chatbots into complex workflows, we approach a tipping point: AI systems operating in autonomous, self-sustaining feedback loops. Any autonomous AI system relies on automated, verifiable reward and feedback mechanisms — evaluation becomes central, but only if the model understands what it is being evaluated on and how results will be used. This work introduces EvalDetectBench, a benchmark measuring LLM "evaluation awareness" — whether a model can recognize the goals, criteria, and potential biases of the evaluations it undergoes.

2. The Memory Trust Gap: Capability-Dependent Failures in Persistent-Memory Agents

Authors: Jundong Hu, Shekar Ramachandran arXiv: 2609.01852

Persistent memory enables personalized agents, but a single stale fact can silently override current authoritative evidence. Testing the Qwen3 series (0.6B/1.7B/4B/8B), the authors find the "memory trust gap" reflects over-trust rather than confusion: on a Benefit set, models at all sizes answer with the stale value 0.92–1.00 of the time; on a Safety set, larger models collapse entirely once stale notes are disguised as current. Mitigations are capability-dependent: exposing metadata helps capable models, but only pre-resolving conflicts restores accuracy on the two smaller checkpoints.

3. Belief-Calibrated Optimization: Optimizing Agents with Explicit World Models

Authors: Yuhan Chen, Zhihua Tian, Mahavir Dabas, et al. arXiv: 2609.01861

LLM agent performance often depends on scaffolds around a frozen model. A common improvement pattern uses an encoder agent as optimizer: it reads current scores and trajectories, iteratively edits source code, and emits a new candidate each round. BCO makes the implicit beliefs behind each edit explicit — written as a persistent in-context document and revised whenever a new candidate is evaluated, effectively serving as a world model of how the environment responds. Across five benchmarks (memory QA, tool-use QA, code-as-action app agents, terminal agents), BCO achieves higher training pass rates than baselines.

4. Epistemic Witch-Attack Resistance: Multiplying AI Agents Without Multiplying Evidence

Author: Marc Bara arXiv: 2609.01873

Multi-agent AI systems improve reasoning by spawning agents and synthesizing reports — but another agent is not another observation. Seemingly independent reports may derive from the same evidence. This paper formalizes the "epistemic witch problem": a report Z is an epistemic witch extension relative to report R when I(Theta; Z | R) = 0. No report-only aggregator can generally distinguish duplication from independent confirmation. Experiments show: with fixed evidence roots, growing reports from 1 to 32 drops naive posterior coverage from 0.940 to 0.263; with fixed report count, growing evidence roots from 1 to 16 fully closes the gap. Collective inference should track evidence ancestry and dependence, not agent or report counts.

5. A Jacobian Lens on Recurrent Transformers: Does the Global Workspace Survive Recurrence?

Authors: Wenlong Wang, Fergal Reid arXiv: 2609.01924

Recent work identifies a verbalizable, causally valid representational band in mid-layers of feedforward transformers — functionally like a global workspace. Does the same arise when depth comes from recurrence rather than stacked layers? Applying a full workspace-analysis suite to Ouro-2.6B (48 layers, 4 loops, depth supervision) and Huginn-0125 (4-layer core, 16 recurrences, trained for latent reasoning), with Qwen3.6-27B as a baseline, the authors find workspaces do form in iterative architectures, but recurrence changes access patterns: Ouro rebuilds workspace content each loop, with linear transport failing across loop boundaries; Huginn carries content forward across all 16 recurrences, while reads/writes/ablations only work within a sliding window of roughly two recurrences.

6. HeadWiseKV: Budgeted Per-Head Cache Residency for Hybrid Long-Context Models

Authors: Renjie Xie, Juncheng Yang, Aoting Hu, et al. arXiv: 2609.02029

Long-context inference retains a growing KV cache during decoding, consuming GPU memory and potentially reducing throughput. HeadWiseKV is a training-free framework that compresses the residual global KV cache while preserving native local, recurrent, and linear paths. It assigns each physical KV head a static multi-level historical window, making cache demand predictable before serving. In fixed-model system studies, it reduces peak sampling device memory by 8.59% at 112K context and extends the maximum verified-success context from 114K to 161K.

7. Monitoring Web Agents Without Internal Signals: Observable Trajectories and Key-Step Supervision

Authors: Sitong Pan, Yipeng Shen, Yilin Lu, et al. arXiv: 2609.02057

When internal uncertainty signals (e.g., token logits) are unavailable, reliable web-agent monitoring is hard. This work studies prefix-level risk prediction from observable trajectory signals: given a progressive prefix, estimate whether execution is on track or heading toward failure. Two representations are derived: macro features summarizing agent–environment behavior and feedback, and micro features measuring consistency of intent, action, and expected state change via repeated black-box queries. On WebArena-Lite and Online Mind2Web, observable-trajectory signals are competitive with internal-signal baselines and support early intervention under fixed false-interruption budgets.

8. When Agents Build Systems: A Case Study in Defects, Detection, and Evaluation Rigor

Author: Phanindra Reddy Madduru arXiv: 2609.01985

As LLM coding agents increasingly perform end-to-end engineering work, we lack empirical characterization of their behavior on system-level requirements: schema design, async orchestration, configuration correctness, and retrieval–filtering trade-offs. This case study catalogs five defects in one agent session implementing a multi-component data system, classified by violated constraint and detection method. HotpotQA evaluation shows that restricting candidates to a graph-identified entity set before re-ranking significantly outperforms unfiltered search across budgets.

---

📋 More Papers

9. AI and Metaethics

Authors: Paul Formosa, Malcolm Ryan · 2609.01793 Explores new metaethical questions raised by AI, including whether machines can hold moral status and ethical frameworks for AI decision-making.

10. When Can Machines Trust Statutes? Certificates of Viability for Digital Law

Author: Marc Lauritsen · 2609.01803 Proposes "certificates of viability" as a verification mechanism for how AI systems validate and trust legal texts.

11. When Does Information Sharing Improve Decentralized Discovery?

2609.01812 Studies conditions under which information sharing helps or hinders decentralized discovery in multi-agent systems.

12. Induction and Inquiry via Probabilistic Reasoning over Linguistic and Program Hypotheses

Authors: Wasu Top Piriyakulkij, Sam Acquaviva, et al. · 2609.01815 A computational model combining neural-driven hypothesis generation with approximate Bayesian updates, reproducing human cognitive effects such as anchoring and garden-pathing in inductive learning and active inquiry.

13. Conversational Data Systems for Stateless LLM APIs: The Hydration Agent Pattern

Author: Joseph Axisa · 2609.01834 An architecture decoupling session persistence from the inference engine, addressing session state management for stateless LLM APIs.

14. SSAKG 2.0: Open-Source Package for Structured Associative Sequence Memory

Authors: Przemysław Stokłosa, Janusz A. Starzyk, Paweł Raif · 2609.01849 Open-source package (Python + C, Apache 2.0) for building structured sequence associative knowledge graphs, reconstructing complete sequences from partially unordered context.

15. Learner Gaps and Measurement Frontiers in Clinical Prediction

Authors: Sayeed Shafayet Chowdhury, et al. · 2609.01909 Distinguishes two causes of clinical-prediction saturation — failure to extract available information vs. a population measurement frontier — validated on three real cohorts (UCI readmission, BRFSS diabetes, NHANES HbA1c).

16. Post-Training Ternarization of Qwen3-4B

Authors: Anirudh Malik, M Sparsh Mehra, Poojith Devan · 2609.01962 End-to-end post-training ternarization of Qwen3-4B: effective 1.641 bits/weight, model compressed from 8.29 GiB to 3.96 GiB, task accuracy falling from 64.5% to 54.7%.

17. StatFormBench: Benchmarking Statistical Problem Formulation in Language Models

Authors: Chen Wang, Junzhe Zhao, et al. · 2609.01982 A benchmark of 1,013 samples across 20 coarse and 85 fine-grained statistical problem categories, evaluating LLMs' ability to formulate statistical problems.

18. ClaimReceipt: Verifying Evidence Sufficiency and Coverage in Agent Evaluation

Authors: Peiying Zhu, Sidi Chang · 2609.01992 A specification addressing whether reported claims can be recomputed from retained evidence (sufficiency) and whether retained records cover the promised set of experiments (coverage).

19. DocHop: A Benchmark for Cross-Domain Multi-Hop Reasoning in Information-Dense Documents

Authors: Zhuoran Yu, Le Thien Phuc Nguyen, et al. · 2609.02059 Tests whether multimodal LLMs can use textual context to determine how chart evidence should be selected, interpreted, and aggregated.

20. MineTRACE: Evidence-Driven Interactive Reasoning for Mineral Exploration

Authors: Yiran Zhang, Jinwen Liu, et al. · 2609.02060 A web-based, evidence-driven exploration system supporting eight commodities (copper, gold, nickel, tungsten, tin, cobalt, tantalum, manganese), using transparent expert trees over multi-source evidence to produce interpretable exploration scores.

---

*Automatically collected 2026-09-04 · Papers.Cool monitoring service*