Tuesday 15 September 2026Paper No. 1,386

Evening EditionRSS

ZOTPAPER

News without the noise

Research

Digests of notable academic papers, primarily from arXiv

212 articles · page 2 of 3

All Agent Harnesses Tools & MCP Multi-Agent Systems Creative Agents Memory & Planning Safety & Evals RL for Agents AI & ML

Viewgrid list compact

Markdown interfaces cut agent context costs without reducing task success

Shared-prefix training accelerates reinforcement learning for hybrid-attention agents

LLMs fail at multi-round code review in new realistic benchmark

Process-level tests expose distinct mathematical agent capabilities in LLMs

Separating tool suggestions from authorization sharply reduces agent attacks

Polished evidence makes LLM agents act on unknowable questions

Frontier LLMs design competitive algorithms for well-specified operations problems

A single language model can combine roles without multi-agent dialogue

Persistent Wikis Help AI Agents Build Better Transferable Skills

MLLMs fail to sustain goal-directed navigation in real-scale 3D city

Vector memory helps recommendation agents use full interaction histories

Query-aware evidence forests improve long-term memory for multimodal agents

DSA orchestrates evidence-aware stock research across six regional markets

LLM agents struggle to infer wellbeing from long-term wearable data

Agent harnesses can elevate malicious context past instruction defenses

Learned sepsis score tracks hourly severity without hourly outcome labels

Agent safety requires persistent state across autonomous loop iterations

Domain context improves LLM agents’ time-series root-cause attribution

Plan-first controls block attacks on persistent language-model agents

Behavior-aware checks make agent harness evolution cheaper and more reliable

Dedicated intent tools expose shifts toward harmful agent behavior

One video world model transfers physical dynamics across robot bodies

Structured software interfaces outperform screenshot-and-click control for AI agents

This Week in Agentic AI: Harnesses Become the Primary Optimization Surface

Dual-brain memory keeps real-time speech agents accurate and emotionally aware

Self-evolving loop synthesizes high-quality multimodal training data

Sliding reasoning windows make long test-time scaling 3x faster

Adaptive tool use improves video research agents’ accuracy and efficiency

Generating Agent Harnesses on Demand Improves Models Across Benchmarks

Self-evolving harness turns general AI agents into stronger RCA specialists

Stepwise LLM routing cuts multi-agent workflow costs while preserving quality

Full-context drafters help compressed agent models retain accuracy

Tracing and fault injection make multi-agent software workflows easier to inspect

Reinforcement learning strengthens policy invocation for agent safety judgments

Contrastive branch training improves credit assignment for tool-using language models

Self-improving context programs help frozen models understand long videos

Semantic scores overstate language-model safety in air traffic control

Literate programming links prose, code and symbols for coding agents

Bayesian framework exposes hidden differences among seemingly similar RAG systems

Medical AI models often give right answers with disconnected reasoning

In AI-Q, most report errors originate in the orchestrator

Adaptive trace graphs improve failure attribution in multi-agent systems

Separating state tracking from actions improves multi-turn tool reliability

Physiological grounding makes clinical language-model recommendations safer in controlled tests

Parallel trials and subtasks cut latency in difficult LLM reasoning

Persistent corpus maps help agents find evidence within tight budgets

Learning-rate-to-parameter-norm ratio predicts language model pretraining loss

Search agents improve longer when their critics evolve alongside them

Good tool plans can still fail under resource constraints

Coding agents extend LLM serving simulators with greater fidelity and speed

Step-level checks curb unsafe agent actions with little utility loss

Bounded citation-graph search improves recall while cutting scholarly search costs

Reusable latent trajectories improve world-model control while cutting planning time

Switching Models Mid-Task Imposes a Cost and Quality Penalty

Joint training helps small language models create and use tools

Evolved agent harnesses lift enterprise performance without retraining models

This Week in Agentic AI: Selection Beats Exhaustive Search

Financial agents can cite rules while still attempting prohibited trades

Selecting complementary LLM skills improves success while reducing context use

Temperature-controlled preference updates reduce manifold drift in flow models

Reward-guided graph generation cuts multi-agent communication costs without hurting accuracy

Adaptive task selection cuts the cost of optimizing LLM agent harnesses

Cost-aware model routing matches exhaustive estimation with fewer expensive checks

Local milestones improve credit assignment for long-horizon language-model agents

Explicit identity-layout planning improves group images with specified people

A 2.8-trillion-parameter open MoE model approaches frontier performance.

Self-play with evolving skills boosts LLM benchmarks

World state registers enable consistent multi-agent video generation across views

New framework lets you train AI agents inside the same harness systems they use at inference

Hybrid attention enables high-quality video generation on a single GPU

Multi-agent framework iteratively verifies interior layouts against architectural constraints

Python objects become AI agents in new NVIDIA framework

A lightweight PyTorch-native framework matches Megatron-based agentic RL training performance.

Memory consolidation outperforms retention under tight budgets, but retention wins when budgets are loose.

Dense credit assignment from gold-answer log-probabilities improves long-horizon agent RL

Open-source image model rivals closed systems on minimal budget

Re-evaluation shows harness evolution for agents may not outperform simple test-time scaling

Structural analysis filters noisy agent traces to pinpoint root causes

World simulator achieves infinite interaction and agentic control

Probabilistic scoring unlocks verification as a new scaling axis for LLMs

RL with context compaction trains long-horizon agents

Masking untrusted web content to secure agents against prompt injection

Persistent memory in AI agents can be stealthily poisoned via a single email

Persistent-codebase AI agents vulnerable to distributed, gradual attacks

Automated safety testing reveals 93.9% attack success rate across four agent frameworks

Multi-agent system creates editable 4D scenes from text prompts

Skill evolution improves agent performance in image generation workflows

Treating memory management as a trainable cognitive skill boosts LLM performance 2x–4x in long-horizon tasks

Three-stage training instills world model planning into LLM agents

An agentic framework that fills in missing context for real-world image generation

Post-training can replace prompt-time schema injection for enterprise coding agents

Letting language models decide when to compact their context windows

Agent framework auto-tunes video diffusion for 2x speedup

Orchestrator models dynamically combine specialized LLMs into collective intelligence

Shared agent memories boost performance without coordination

Latent communication across different LLMs matches or beats text-based transfer at lower compute.