Today’s AI & Tech Briefing (June 19, 2026)

Today’s selection of 8 noteworthy AI/ML papers from arXiv, covering multi-agent systems, knowledge conflicts in LLMs, educational AI, multimodal reasoning, cybersecurity, compiler optimization, retrieval systems, and reinforcement learning under missing data.

1. Contagion Networks: Evaluator Bias Propagation in Multi-Agent LLM Systems

Authors: Zewen Liu | Categories: cs.LG, cs.AI, cs.MA Link: arxiv.org/abs/2606.20493v1

Introduces a formal framework called Contagion Networks to measure how evaluator biases spread across interacting LLM agents. In a controlled 3-agent experiment with DeepSeek-chat, biases consistently propagated between agents (gamma in [0.157, 0.352]), even within the same model. Increasing the evaluator committee size from 1 to 3 reduced effective contagion by 72.4%.

Takeaway: As multi-agent LLM systems become more prevalent, this work provides both a crucial diagnostic tool and a practical mitigation strategy for the often-overlooked problem of bias amplification through agent interactions.

2. Navigating Unreliable Parametric and Contextual Knowledge: Explicit Knowledge Conflict Resolution for LLM Inference

Authors: Huang Peng, Jiuyang Tang, Weixin Zeng, Hao Xu, Xiang Zhao | Categories: cs.AI Link: arxiv.org/abs/2606.20245v1

Proposes MACR, a novel framework that moves beyond binary choice paradigms (trusting either parametric or contextual knowledge) to actively resolve inconsistencies using a multi-agent reasoning approach. The framework uses a modified semantic entropy measure for confidence estimation and three specialized agents to induce rules, analyze conflicts, and resolve inconsistencies across all available contexts.

Takeaway: This addresses a fundamental blind spot in existing approaches—that both internal knowledge and external context can be unreliable—and offers interpretable conflict resolution, which is critical for deploying LLMs in high-stakes applications.

3. PsyScore: A Psychometrically-Aware Framework for Trait-Adaptive Essay Scoring and ZPD-Scaffolded Feedback

Authors: Wei Xia, Jin Wu, Haoran Shi, Xiangyu Wang, Chanjin Zheng | Categories: cs.CL Link: arxiv.org/abs/2606.20287v1

Integrates diagnostic assessment with instructional scaffolding through a shared latent ability representation, combining neural IRT scoring with zone-of-proximal-development-adaptive feedback generation. On the ASAP++ dataset, PsyScore achieves competitive scoring while providing more pedagogically aligned, proficiency-level-specific feedback.

Takeaway: A rare paper that genuinely bridges psychometrics and modern AI, moving beyond the typical separation between scoring and feedback in automated essay evaluation systems.

4. Evaluating and Enhancing Negation Comprehension in Remote Sensing MLLMs

Authors: Haochen Han, Jue Wang, Alex Jinpeng Wang, Fangming Liu | Categories: cs.CV, cs.AI Link: arxiv.org/abs/2606.20177v1

Introduces RS-Neg, the first benchmark to evaluate negation understanding in remote sensing MLLMs across region-level to scene-level tasks. Reveals that advanced RS MLLMs struggle with negation, exhibiting hallucinations and substantial performance degradation. Proposes NeFo, a test-time learning method that uses about 5% unlabeled test samples to significantly improve negation understanding.

Takeaway: Negation comprehension is a critical failure mode for real-world deployment of vision-language models, especially in emergency scenarios like identifying non-flooded evacuation routes—this work identifies and begins to close that gap.

5. Multi-View Decompilation for LLM-Based Malware Classification

Authors: Bercan Turkmen, Vyas Raina | Categories: cs.CR, cs.AI Link: arxiv.org/abs/2606.20436v1

Shows that providing decompiled views from both Ghidra and RetDec to LLMs improves malicious-class F1 scores by increasing recall on malicious samples. The key insight is that different decompilers make partially different errors, providing complementary evidence that a single-view pipeline would miss.

Takeaway: A simple, training-free method that meaningfully improves LLM-assisted malware triage—practical cybersecurity work that acknowledges the inherent lossiness of decompilation tools.

6. AutoPass: Evidence-Guided LLM Agents for Compiler Performance Tuning

Authors: Zepeng Li, Jie Ren, Zhanyong Tang, Jie Zheng, Zheng Wang | Categories: cs.SE, cs.AI Link: arxiv.org/abs/2606.20373v1

Opens up the compiler internals to LLM agents, enabling them to query optimization states and analyze intermediate representations to guide performance tuning. Achieves geometric-mean speedups of 1.043x and 1.117x over LLVM -O3 on x86-64 and ARM64 respectively, outperforming both expert-tuned heuristics and classical autotuning methods.

Takeaway: A compelling demonstration that LLMs can go beyond black-box code generation to perform white-box optimization of complex systems, with immediate practical value for compiler engineering.

7. ELVA: Exploring Ranking-Driven Universal Multimodal Retrieval

Authors: Yuhan Liu, Pei Fu, Hang Li, Yukun Qi, Chao Jiang et al. | Categories: cs.IR, cs.AI Link: arxiv.org/abs/2606.20280v1

Addresses “grain blindness” in multimodal retrieval—where models fail to capture grain-level information in queries—by extending reinforcement learning with verifiable rewards to retrieval tasks. Achieves state-of-the-art results across standard benchmarks, with a notable 13.1% improvement on their new multi-grain query benchmark MRBench.

Takeaway: The grain blindness problem is a subtle but important limitation of contrastive learning for retrieval; this RL-based approach offers a principled solution without requiring explicit ranking labels.

8. Off-Policy Evaluation for Missingness-Aware Policies in MDPs with Rewards Missing Not at Random

Authors: Ziheng Wei, Annie Qu, Rui Miao | Categories: stat.ML, cs.LG Link: arxiv.org/abs/2606.20206v1

Addresses the critical problem of rewards missing not at random in offline RL, which introduces selection bias even after conditioning on states and actions. Proposes a fitted-Q-evaluation-style estimator that uses future states as shadow variables and a bridge function to recover conditional mean rewards without explicitly modeling the missingness mechanism.

Takeaway: Missing reward data is a pervasive real-world problem in healthcare and marketing applications of RL; this work provides rigorous theoretical grounding and practical estimation methods for a previously understudied setting.

This content was generated with AI assistance. Paper information sourced from arXiv.