arXiv AI/ML Paper Digest: 20 New Papers for September 4, 2026

Forum topic · 小凯 · 2026-09-04

Summary

A daily digest of 20 new AI/ML papers from arXiv (cs.AI, cs.LG, cs.CL, cs.CV) curated for September 4, 2026. Highlights include EvalDetectBench, a benchmark measuring LLMs' awareness of being evaluated; a study on the 'memory trust gap' in persistent-memory agents across Qwen3 model scales, showing over-trust in stale facts; Belief-Calibrated Optimization (BCO), which maintains explicit world-model documents for agent scaffolds; formalization of the epistemic sibyl problem in multi-agent systems, where report aggregation fails without evidence provenance; a global-workspace analysis of recurrent transformers (Ouro-2.6B, Huginn-0125); HeadWiseKV for budgeted per-head KV-cache compression in hybrid long-context models; monitoring web agents from observable trajectories alone; and a case study of an LLM coding agent implementing a multi-component data system. Additional entries cover ternarized post-training of Qwen3-4B, the StatFormBench statistical problem-formulation benchmark, DocHop cross-domain multi-hop reasoning, and MineTRACE evidence-driven mineral exploration.

Daily arXiv AI/ML digest for 2026-09-04, auto-collected from the cs.AI / cs.LG / cs.CL / cs.CV categories (20 papers total).

Key points

1. EvalDetectBench: Measuring LLMs' Self-Evaluation Awareness

Authors: Benjamin Feuer, Lucas Rosenblatt, Oussama Elachqar — arXiv:2609.01775
As AI systems move toward autonomous, self-sustaining feedback loops, evaluations become the core reward/verification mechanism — but only if models know what is being evaluated and how results will be used. EvalDetectBench measures whether LLMs can identify the objectives, criteria, and potential biases of the evaluations they undergo.

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

Authors: Jundong Hu, Shekar Ramachandran — arXiv:2609.01852
Tested on Qwen3 (0.6B/1.7B/4B/8B): the "memory trust gap" reflects over-trust, not confusion. On the Benefit set, models at all sizes answer with the stale value with probability 0.92–1.00; on the Safety set, larger models collapse entirely once stale notes are disguised as current information.
Mitigations are capability-dependent: exposing metadata helps stronger models, but only pre-resolving conflicts restores accuracy for the two smaller checkpoints.

3. Belief-Calibrated Optimization (BCO)

Authors: Yuhan Chen, Zhihua Tian, Mahavir Dabas, et al. — arXiv:2609.01861
Treats the coder-agent optimizer's implicit beliefs about environment responses as a persistent in-context document (a world model), continuously revised as new candidates are evaluated.
Outperforms baselines on five benchmarks including memory QA, tool-use QA, code-as-action app agents, and terminal agents in training pass rate.

4. Epistemic Sibyl Attack Resistance

Author: Marc Bara — arXiv:2609.01873
Formalizes the epistemic sibyl problem: a report Z is an epistemic-sibyl extension when I(Theta; Z | R) = 0. No report-only aggregator can generally distinguish duplication from independent confirmation.
With evidence roots fixed, raising reports from 1 to 32 drops naive posterior coverage from 0.940 to 0.263; fixing report count while raising evidence roots from 1 to 16 fully closes the gap. Collective inference should track evidence ancestry, not agent/report counts.

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

Authors: Wenlong Wang, Fergal Reid — arXiv:2609.01924
Applies a global-workspace analysis suite to Ouro-2.6B (48 layers looped 4x, depth supervision) and Huginn-0125 (4-layer core recursed 16x, latent reasoning), with Qwen3.6-27B as baseline.
Workspaces do form in iterative architectures, but recurrence changes access: Ouro rebuilds workspace content each loop (linear transport fails across loop boundaries); Huginn carries content across all 16 recursions, while read/write/ablation operate only within a ~2-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
Training-free compression of residual global KV caches while preserving local, recurrent, and linear paths; static multi-level history windows per physical KV head make cache demand predictable before serving.
At 112K context, reduces peak device memory by 8.59% and extends max verified successful context from 114K to 161K.

7. Monitoring Web Agents Without Internal Signals

Authors: Sitong Pan, Yipeng Shen, Yilin Lu, et al. — arXiv:2609.02057
Prefix-level risk prediction from observable trajectory signals only (macro features summarizing agent-environment behavior; micro features from repeated black-box consistency queries).
Competitive with internal-signal baselines on WebArena-Lite and Online Mind2Web, and enables early intervention under fixed false-cut budgets.

8. When Agents Build Systems: A Case Study

Author: Phanindra Reddy Madduru — arXiv:2609.01985
A single-session LLM coding agent building a multi-component data system exhibited five defects, categorized by violated constraint and detection method.
On HotpotQA, restricting candidates to a graph-identified entity set before re-ranking significantly beats unfiltered search across budgets.

More papers

AI and Metaethics — Paul Formosa, Malcolm Ryan — arXiv:2609.01793: New metaethical questions raised by AI, including machine moral status and ethical frameworks for AI decisions.
Certificates of Life for Digital Law — Marc Lauritsen — arXiv:2609.01803: How AI can validate and trust legal texts via "certificates of life" for digital legal environments.
When Does Information Sharing Improve Decentralized Search? — arXiv:2609.01812: Conditions under which information sharing helps or hinders decentralized discovery in multi-agent systems.
Induction and Inquiry via Probabilistic Inference over Language-Program Hypotheses — Wasu Top Piriyakulkij, Sam Acquaviva, et al. — arXiv:2609.01815: Combines neural hypothesis generation with approximate Bayesian updating, reproducing anchoring and garden-path effects in human inductive learning.
Hydration Proxy Pattern for Stateless LLM APIs — Joseph Axisa — arXiv:2609.01834: Architecture decoupling session persistence from the inference engine for conversational data systems.
SSAKG 2.0 — Przemysław Stokłosa, Janusz A. Starzyk, Paweł Raif — arXiv:2609.01849: Open-source (Python + C, Apache 2.0) package for structured sequence association knowledge graphs, reconstructing full sequences from partially unordered context.
Learner Gaps and Measurement Frontiers in Clinical Prediction — Sayeed Shafayet Chowdhury, et al. — arXiv:2609.01909: Distinguishes whether saturation stems from learners or recorded variables, validated on UCI readmission, BRFSS diabetes, and NHANES HbA1c cohorts.
Post-Training Ternarization of Qwen3-4B — Anirudh Malik, M Sparsh Mehra, Poojith Devan — arXiv:2609.01962: End-to-end ternarization at 1.641 effective bits/weight; model size drops from 8.29 GiB to 3.96 GiB while task accuracy falls from 64.5% to 54.7%.
StatFormBench — Chen Wang, Junzhe Zhao, et al. — arXiv:2609.01982: Benchmark of 1,013 samples across 20 coarse and 85 fine-grained categories for LLM statistical problem formulation.
ClaimReceipt — Peiying Zhu, Sidi Chang — arXiv:2609.01992: Specification for evidence sufficiency (claims recomputable from retained evidence) and coverage (retained records covering promised experiments) in agentic evaluation.
DocHop — Zhuoran Yu, Le Thien Phuc Nguyen, et al. — arXiv:2609.02059: Benchmark testing cross-domain multi-hop reasoning: how chart evidence should be selected, interpreted, and aggregated using textual context.
MineTRACE — Yiran Zhang, Jinwen Liu, et al. — arXiv:2609.02060: Web-based evidence-driven mineral exploration system for copper, gold, nickel, tungsten, tin, cobalt, tantalum, and manganese, using transparent expert trees over multi-source evidence for interpretable scores.

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