11 articles — 🟡 7 important , 🟢 4 interesting
A paper accepted at ICML 2026 systematically tests LLM tool-use agents under environment shifts across four levels — Perception, Interaction, Reasoning, and Internalization. Findings: both SFT and RL training show significant degradation under modest distribution shifts, and controlled benchmark accuracy does not predict real-world robustness. The proposed PAFT (Perturbation-Augmented Fine-Tuning) offers mitigation.
The SEA (Self-Evolving Agents with Anytime-Valid Certificates) architecture allows agents to update their own parameters while retaining formal learning-theoretic guarantees. Five verification mechanisms and auditable certificates approve or block each self-modification in real time, achieving +4 to +5 solved instances on SWE-bench Verified with strong base models.
GitHub has launched a public preview of session streaming for Copilot agents, giving enterprise organizations direct visibility into the prompts, responses, and tool calls that agents execute across all Copilot clients.
Stanford researchers developed AutoMem — a system with two optimization loops that automatically learns how to organize and use memory, without human annotation, achieving 2–4× improvement over baselines.
The AWS SageMaker AI guide for multi-turn reinforcement learning places reward function quality and evaluation independence ahead of algorithmic choice. Dense rewards prevent variance collapse, and reward hacking occurs when the agent optimizes the metric without solving the actual task. On the SOP-Bench benchmark, correctly configured training achieved a 13 percent better Task Success Rate and around 16 percent better per-field accuracy.
GitHub released two enterprise updates for Copilot on the same day: the CLI in GitHub Actions no longer requires a personal access token, and cost centers now support AI credit pools with automatic limits that protect teams from one another.
LangChain describes how coding agents generate uncontrolled costs due to tool fragmentation and a tokenmaxxing mindset, and proposes a four-phase approach via the LangSmith platform covering visibility, normalization, optimization, and governance.
Researchers discovered why jailbreaks succeed at the level of internal model representations and developed the HARC fine-tuning method, which explicitly couples 'harmfulness and refusal directions' — achieving the strongest robustness-capability-usability trade-off among six tested methods.
Amazon Bedrock foundation models detect AI-generated phishing by analyzing email content behavior patterns, not surface spam signals. A five-stage pipeline combines authentication checks, AI analysis, and multi-factor risk scoring on a scale of zero to one hundred. A continuous-learning system in a five-phase feedback loop improves detection accuracy with experience.
