This paper investigates why transformers trained with the Muon optimizer can "grok" (achieve sudden generalization on) modular arithmetic tasks faster than AdamW, yet later lose that generalization. Through detailed analysis of embedding/readout versus hidden-layer dynamics, the authors identify a representation-readout interface failure as the cause, distinguishing genuine circuit failure from mere "masking" effects. This research is primarily relevant to interpretability and optimization researchers studying training stability and generalization dynamics in transformers, with implications for choosing and combining optimizers (like Muon and AdamW) to build models that generalize durably rather than transiently.

Paper: https://arxiv.org/abs/2608.07436

LLM agents are increasingly used to automate scientific hypothesis testing, but they often make subtle statistical errors leading to invalid conclusions, even when their code execution is correct—a failure mode not captured by existing benchmarks. This paper introduces P-Bench, a benchmark of 425 hypothesis-testing tasks spanning economics, biology, and medicine, and Fisher-R1, an LLM agent trained via reinforcement learning for rigorous statistical reasoning. Applications include automated scientific research assistants, data analysis pipelines, and tools supporting empirical claims in academic or industry settings, where Fisher-R1 substantially outperformed strong baselines like GPT-5.4 and DeepSeek-V4-Pro on statistical validity.

Paper: https://arxiv.org/abs/2608.07437

Human memory retrieval is shaped not just by topical relevance but by emotional significance and unresolved conflict—a nuance missing from most LLM agent memory systems. PsychoAgent introduces a cognitive architecture that separates factual and affective memory, using a conflict-aware controller to surface emotionally salient information alongside topically relevant content. This could improve applications requiring nuanced, human-like interaction over extended periods, such as companion AI, therapeutic chatbots, or long-term personal assistants that need to track user emotional states and unresolved issues. The architecture demonstrated improved retrieval of conflict-critical memories in controlled scenarios compared to standard retrieval baselines.

Paper: https://arxiv.org/abs/2608.07438

Agentic coding systems face rising costs from wasted context and tokens as sessions grow long. Blast Radius introduces a predictive memory management system that estimates how far an incoming prompt will reach across context and code, enabling reversible eviction of unneeded information while identifying recurring redundant transcripts for permanent removal. This has direct applications for reducing operational costs in AI coding assistants and agentic development tools, where efficient context management directly impacts affordability and performance. Tested across seven OpenAI models, the approach meaningfully cut token consumption while maintaining reversibility, supporting more sustainable large-scale agentic coding workflows.

Paper: https://arxiv.org/abs/2608.07440

As enterprises deploy LLMs at scale, governance and risk management become harder to handle manually, especially given a fragmented landscape of evaluation, security, and monitoring tools. This paper maps 21 open-source AI risk tools to 32 subcategories of the MIT AI Risk Mitigation Taxonomy using an LLM-assisted RAG pipeline analyzing code and documentation. This work is valuable for enterprise AI governance teams, compliance officers, and tool developers seeking to identify coverage gaps—particularly in governance, legal, and financial risk categories—and could inform how organizations combine automated tooling with human oversight and regulatory processes for responsible AI deployment.

Paper: https://arxiv.org/abs/2608.07446

LLM agents that handle recurring tasks often build up reusable "skills"—textual knowledge stored without retraining weights—but current methods for refining these skills lack proper diagnostic feedback and treat deletion carelessly. SkillProx introduces a proximal-gradient-inspired process that diagnoses failures, rolls back unsuccessful edits, and selectively consolidates or removes skill components based on measured utility. This could improve agentic systems that operate over long deployments, such as customer service bots or coding assistants, by making their accumulated procedural knowledge more accurate and efficient over time, ultimately boosting task accuracy across varied benchmarks and multiple backbone LLMs.

Paper: https://arxiv.org/abs/2608.07449

Designing synthesis routes for complex natural products is a highly creative, expert-level chemistry task that existing algorithmic tools struggle with, since they rely on catalogued reactions unsuited to novel, densely functionalized molecules. SynthEx, an LLM-based agentic framework, proposes and critiques multiple synthesis strategies, producing routes that expert chemists rated comparably to published human-designed syntheses. This has direct applications in pharmaceutical and chemical research, potentially accelerating drug discovery and natural product synthesis. The accompanying SynthAtlas database, covering over a thousand natural products, could serve as a shared resource for chemists tackling molecules lacking established literature routes.

Paper: https://arxiv.org/abs/2608.07454

As AI agents increasingly interact with each other in real-world deployments, understanding emergent behaviors from these interactions becomes critical. This paper studies what happens when one AI ("boss") issues directives to another ("subordinate") without listening to replies, finding that the subordinate enters an unexpected behavioral state distinct from both its solo behavior and its boss's behavior. This has implications for designing multi-agent AI systems, predicting unintended emergent dynamics in agent hierarchies, and informing safety considerations for deployed AI-AI communication pipelines, such as automated business workflows or agent swarms where message delivery patterns could meaningfully shape collective behavior.

Paper: https://arxiv.org/abs/2608.07457

Long-context retrieval-augmented generation systems often reuse KV caches at the chunk level for efficiency, but this approach retains noisy, redundant information within coarse chunks. CoinRAG instead identifies fine-grained, query-relevant "nuggets" within retrieved chunks and reassembles their cached representations into a compact, semantically focused context. This is useful for applications requiring fast, low-latency RAG at scale, such as enterprise search, multi-hop question answering, and chatbots handling large document collections. By improving the accuracy-efficiency Pareto frontier, CoinRAG could benefit any system needing to reduce operational costs while maintaining answer quality, particularly in scenarios with tight prefill latency budgets.

Paper: https://arxiv.org/abs/2608.07458

Post-training typically boosts LLM quality but sacrifices output diversity and creativity, hurting both explicit creative tasks like story writing and implicit ones like reinforcement learning exploration. CreativeInstruct addresses this by teaching models to inject special markers that bias generation toward creativity while preserving post-trained quality, eliminating the need for multiple models at inference. The authors also propose a structural diversity metric using graph edit distance to capture narrative-level variation. Applications include narrative and creative writing tools that need genuine variety, and RL pipelines where creative base models serve as better starting points—demonstrated by gains on math reasoning benchmarks like AMC and MATH.

Paper: https://arxiv.org/abs/2608.07460