Agents and Memory Safety, Vision-Language Action Models, and Reasoning Under Uncertainty

Today’s selection of 8 noteworthy AI/ML papers from arXiv spans AI safety mechanics, uncertainty quantification in LLMs, efficient critic training for RL, open-world video agents, domain-specific RAG pipelines, novel attack vectors on agent memory, robot learning via intent distillation, and a systematic study of agent decomposition strategies.

1. Mitigating Reasoning-Induced Misalignment via Safety-Direction Penalty

Authors: Yipeng Zhao, Qishun Yang, Shenzhe Zhu, Shu Yang, Di Wang | Categories: cs.AI, cs.CL Link: arxiv.org/abs/2608.23497

This paper diagnoses Reasoning-Induced Misalignment (RIM), where fine-tuning on benign reasoning data (math, code, chain-of-thought) inadvertently degrades LLM safety. The authors trace this to a geometric coupling between reasoning and safety directions in activation space and propose the Safety-Direction Penalty (SDP) to penalize displacement along the safety axis during fine-tuning. SDP restores safety on Qwen2.5 models without sacrificing benchmark reasoning performance.

Takeaway: Offers both a mechanistic explanation (direction coupling in representation space) and a practical training-time fix, moving beyond mere observation of the problem to actionable mitigation.

2. Credal Large Language Models for Semantic Commitment under Uncertainty

Authors: Shireen Kudukkil Manchingal, Sofiia Nikolenko, Fabio Cuzzolin | Categories: cs.CL, cs.AI, cs.LG, stat.ML Link: arxiv.org/abs/2608.23244

This work introduces Credal LLMs (CLLMs), which use an ensemble of LoRA adapters to produce a credal set of predictive distributions rather than a single softmax, thereby distinguishing epistemic ignorance from semantic ambiguity. The authors derive token-level (CTC) and semantic-level (SCC) commitment scores, achieving state-of-the-art hallucination detection and near-perfect selective prediction accuracy (99.0% on OpenBookQA at 80% coverage).

Takeaway: A principled way to surface model uncertainty without extra generation cost, with strong empirical results across multiple backbones and benchmarks.

3. How to Train a Critic Stably and Efficiently

Authors: Penghui Qi, Xiangxin Zhou, Wee Sun Lee | Categories: cs.LG, cs.AI, cs.CL Link: arxiv.org/abs/2608.23566

The authors analyze instabilities in critic-based RL for LLMs and propose Best-Practice Critic Optimization (BPCO), a recipe combining DPPO, bounded value predictions, Monte Carlo targets, and length-adaptive GAE. BPCO also conditions the critic on reward-defining information (e.g., reference answers) hidden from the policy, enabling it to match or exceed group-based baselines like GRPO while sampling only one response per prompt.

Takeaway: A practical, well-controlled recipe that reinvigorates critic-based methods as a reliable, sample-efficient alternative to group-relative advantage estimation.

4. Thinking Beyond Videos: Unifying Video Reasoning and Deep Research for Open-World Video Agents

Authors: Wenqi Liu, Shijie Ma, Yunxiao Wang, Meng Liu, Qile Su et al. | Categories: cs.CV, cs.AI Link: arxiv.org/abs/2608.23329

VideoRover is a unified framework that iteratively coordinates video cropping, multimodal search, and web browsing to solve open-world video questions requiring external knowledge. Built with 26K SFT trajectories and 3K RL instances, VideoRover-8B-RL matches proprietary models in direct-answer settings and outperforms larger open-source models with the same tool suite on the new VideoRover-Bench.

Takeaway: Demonstrates the payoff of tightly coupling active video grounding with external retrieval, backed by a novel benchmark and training pipeline.

5. Retrieval-Augmented Classification of Environmental Mitigations in Hydropower Licensing Documents

Authors: Hong-Jun Yoon, Tom Ruggles, Joanna Lee, Debjani Singh | Categories: cs.IR, cs.AI Link: arxiv.org/abs/2608.23241

This paper tackles multi-label classification of environmental mitigations across a 135-category taxonomy with severe label scarcity, where 40 categories have zero training examples. A hybrid BERT-detection + RAG-classification pipeline achieves Micro F1 of 0.524, outperforming both BERT-only (0.477) and RAG-only (0.416) approaches by exploiting BERT’s recall and RAG’s zero-shot coverage.

Takeaway: A compelling real-world use case showing how retrieval-augmented reasoning enables zero-shot generalization in domain-specific, high-stakes document classification.

6. InjecMEM: Memory Injection Attack on LLM Agent Memory Systems

Authors: Hanling Tian, Gengyu Zhang, Zeyang Sha, Jingying Wang, Yuhang Liu et al. | Categories: cs.CR, cs.AI Link: arxiv.org/abs/2608.23471

InjecMEM is a single-interaction attack on LLM agent memory systems that steers future responses toward pre-specified outputs via a retriever-agnostic anchor and an adversarially optimized command. The attack remains effective under memory drift and across backbones, while leaving non-target queries unaffected, highlighting a critical new vulnerability surface.

Takeaway: An important and timely security contribution—agent memory is a growing attack surface, and this paper provides a reproducible framework for studying and hardening it.

7. Act with Intent: Distilling Behavior Intent for Vision-Language-Action Models

Authors: Sangoh Lee, Sangwoo Mo, Wook-Shin Han | Categories: cs.RO, cs.AI, cs.CV Link: arxiv.org/abs/2608.23478

Intention Distillation (INDI) trains VLA action decoders to recover a frozen teacher VLM’s multimodal intent representation from standard inputs, using it to organize action prediction beyond behavior cloning. INDI boosts GR00T-N1.7 from 64.3% to 84.7% on SimplerEnv-Bridge and shows consistent gains on real-world long-horizon tasks (+12.0 pp).

Takeaway: Provides strong evidence that action decoders benefit from explicitly modeling the semantic objective of behavior, not just the motor commands.

8. Right-Sizing LLM-Agent Decomposition in VAT Determination: A Pilot Controlled Sweep

Authors: Pedro Santos | Categories: cs.MA, cs.AI, cs.SE Link: arxiv.org/abs/2608.23395

This preregistered pilot systematically varies subtask-to-worker assignment across four orchestrated configurations (one wide worker to five narrow ones) on a bounded VAT-determination domain with oracle labels. Intermediate configurations lead on accuracy (0.830 vs. endpoints at 0.720 and 0.770), but the pre-stated falsification bar is missed; fault-injection arms reveal that wide-scope restart over-recovers while schema-conforming hallucinations degrade all configurations.

Takeaway: A rigorous, reproducible methodology for studying agent decomposition trade-offs, with the intriguing finding that mid-granularity often wins—though evidence remains inconclusive at pilot scale.

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