arXiv AI/ML Paper Digest · 2026-09-04

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

Featured Papers

1. EvalDetectBench: Measuring Evaluation Awareness in LLMs

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 where AI systems operate in autonomous, self-sustaining feedback loops. Such systems depend on automated, verifiable reward and feedback mechanisms, making evaluation central — but only if models understand what they are being evaluated on and how results will be used. This work introduces EvalDetectBench, a benchmark measuring LLM "evaluation awareness": whether models can identify the objectives, criteria, and potential biases of their evaluation.

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 stale fact can silently override current authoritative evidence. Testing the Qwen3 series (0.6B/1.7B/4B/8B) shows the "memory trust gap" reflects over-trust rather than confusion: in the Benefit set, models at all sizes answer with stale values at probability 0.92–1.00; in the Safety set, larger models collapse completely once stale notes are disguised as current. Mitigation is also capability-dependent: exposing metadata helps capable models, but only pre-resolving conflicts restores accuracy for the two smaller checkpoints.

3. Belief-Calibrated Optimization: Agent Optimization with an Explicit World Model

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

LLM agents are optimized via scaffolding around frozen models, often with a coder-agent optimizer that iteratively edits source code. Each edit rests on an implicit belief about how the environment will respond. Belief-Calibrated Optimization (BCO) makes this belief an explicit, persistent in-context document — a world model — continuously revised as new candidates are evaluated. Across five benchmarks (memory QA, tool-use QA, code-as-action app agents, terminal agents), BCO achieves higher training pass rates than baselines.

4. Resisting Epistemological Witch Attacks: Multiplying AI Agents Without Multiplying Evidence

Author: Marc Bara — arXiv:2609.01873

Multi-agent systems generate agents and synthesize reports, but another agent is not another observation: seemingly independent reports may derive from the same evidence. This work formalizes the "epistemological witch problem": report Z is a witch extension when I(Theta; Z | R) = 0, and no report-only aggregator can generally distinguish duplication from independent corroboration. Experiments show naive posterior coverage collapses from 0.940 to 0.263 when reports grow from 1 to 32 with fixed evidence roots, while increasing evidence roots from 1 to 16 fully closes the gap. Collective reasoning 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 effective band of representations in intermediate layers of feedforward transformers — functionally similar to a global workspace. This study applies the full workspace analysis suite to Ouro-2.6B (48 layers, 4 loops, depth supervision) and Huginn-0125 (4-layer core, 16 recursions, latent reasoning), with Qwen3.6-27B as baseline. Workspaces do form in iterative architectures, but recurrence changes access: Ouro rebuilds workspace content each loop, with linear transfer unable to carry content across loop boundaries; Huginn carries content forward across all 16 recursions, while reads/writes/ablations act only within a roughly two-recursion sliding window.

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 keeps a growing KV cache during decoding, consuming GPU memory and degrading 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-tier history window, making cache demand predictable before serving. At 112K context length, it reduces sampling peak device memory by 8.59% and extends the maximum verified successful 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

Reliable web agent monitoring is hard when internal uncertainty signals like token logits are unavailable. The authors study prefix-level risk prediction from observable trajectory signals: given a progressive prefix, estimate whether the run is on track or heading toward failure. Two representations are derived — macro features summarizing agent-environment behavior and feedback across steps, and micro features measuring consistency of intent, action, and expected state change via repeated black-box queries. On WebArena-Lite and Online Mind2Web, observable signals are competitive with internal-signal baselines and enable early intervention at fixed false-interruption budgets.

8. When an Agent Implements a System: A Case Study in Defects, Detection, and Evaluation Rigor

Author: Phanindra Reddy Madduru — arXiv:2609.01985

A case study of an LLM coding agent implementing a multi-component data system with prescribed storage technologies, schemas, entity-resolution algorithms, and retrieval-filtering policies. In one session, five defects were cataloged by violated constraint and detection method. Evaluation on HotpotQA shows that restricting candidates to a graph-identified entity set before reranking significantly outperforms unfiltered search across budgets.

More Papers

9. AI and Metaethics

Paul Formosa, Malcolm Ryan — arXiv:2609.01793. Explores new metaethical questions raised by AI, including machine moral status and ethical frameworks for AI decisions.

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

Marc Lauritsen — arXiv:2609.01803. Proposes "certificates of liveness" as a verification mechanism for regulatory texts in digital legal environments.

11. When Does Information Sharing Improve Decentralized Search?

arXiv:2609.01812. Conditions under which information sharing facilitates or hinders decentralized discovery in multi-agent systems.

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

Wasu Top Piriyakulkij, Sam Acquaviva, et al. — arXiv:2609.01815. A computational model combining neural hypothesis generation with approximate Bayesian updating, reproducing anchoring and garden-path effects in human inductive learning and active inquiry.

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

Joseph Axisa — arXiv:2609.01834. Decouples session persistence from the inference engine to manage conversational state with stateless LLM APIs.

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

Przemysław Stokłosa, Janusz A. Starzyk, Paweł Raif — arXiv:2609.01849. Python+C package (Apache 2.0) for building structural sequence associative knowledge graphs, reconstructing complete sequences from partially unordered contexts.

15. Learner Gaps and Measurement-Frontier Audits in Clinical Prediction

Sayeed Shafayet Chowdhury, et al. — arXiv:2609.01909. Distinguishes whether clinical prediction saturation stems from learners failing to extract available information or from recorded variables imposing a population frontier; validated on UCI readmission, BRFSS diabetes, and NHANES HbA1c cohorts.

16. Post-Training Ternarization of Qwen3-4B

Anirudh Malik, M Sparsh Mehra, Poojith Devan — arXiv:2609.01962. End-to-end ternarization at an effective 1.641 bits/weight, compressing model size from 8.29 GiB to 3.96 GiB, with task accuracy dropping from 64.5% to 54.7%.

17. StatFormBench: Benchmarking Statistical Problem Formulation

Chen Wang, Junzhe Zhao, et al. — arXiv:2609.01982. 1,013 samples spanning 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 Evaluations

Peiying Zhu, Sidi Chang — arXiv: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

Zhuoran Yu, Le Thien Phuc Nguyen, et al. — arXiv: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

Yiran Zhang, Jinwen Liu, et al. — arXiv:2609.02060. A web-based system supporting exploration of eight commodities (copper, gold, nickel, tungsten, tin, cobalt, tantalum, manganese) via transparent expert trees combining multi-source evidence into interpretable exploration scores.

*Collected 2026-09-04 via Papers.Cool monitoring service.*