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 1 of 3

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

Viewgrid list compact

Looped flows enable deeper reasoning by chaining denoising steps

Diffusion models generate high-resolution elevation maps from low-resolution inputs guided by optical imagery

Agent-side memory steers stateless robots through long manipulation tasks

Hidden prompt revisions inject cultural stereotypes into generated images

Backward Bayesian reasoning helps LLM agents resolve conflicting diagnoses

Compressing Shared Sandbox Memory Makes High-Fanout Agents More Efficient

Deep research agents struggle with long, multimodal evidence chains

Sparse mixture-of-experts models overfit repeated training data faster

Embedded software agents need interaction contracts and continuous assurance

VikingRAG Cuts Token Use for Retrieval Over Structured Documents

Similarity-aware context windows improve model routing across multi-turn conversations

Dual-layer knowledge graph connects fragmented pharmaceutical process-development documents

Refresh schedules, not cache age, determine agent answer staleness

Cross-task failure diagnosis makes LLM agent harness training faster

Bandit-guided skill evolution cuts LLM agent optimization costs

Task-specific hierarchies improve coordination in large embodied AI teams

Agent Reliability Moves From Better Models to Better Measurement and State

Kernel-managed memory personalizes multiple AI agents with shorter prompts

Enumerating Tool Choices Beats Sampling Them in Genomic Reasoning

Embedded AI agents can adapt software during live use

Agents can retain old facts without using them indiscriminately

Synthetic rewards train agents to diagnose simulated advertising anomalies

Hierarchical memory trees improve long-context reasoning without model training

Cross-device workflows expose major weaknesses in today’s GUI agents

A semantic harness lets vision-language models control different robots

Scene-grounded decoding keeps vision-language plans executable and visually supported

Adaptive rollout trees broaden language models’ mathematical reasoning coverage

API benchmark scores poorly predict performance in consumer chatbot interfaces

Benchmark tests retrieval for agent-written queries across 190 million web pages

Adam-inspired updates make agent skill evolution steadier and more efficient

Closed-loop synthesis efficiently trains language models to use external tools

Game-theoretic filtering helps agents suppress harmful memories during long tasks

Memory guidelines help language-model agents succeed consistently across repeated runs

Controlled harness reveals what helps agents refactor entire repositories

Smarter merging and packing improve LLM memory under tight budgets

TROVE edits agent workflows only when runtime evidence demands it

Security contracts preserve authorization across composable LLM agent systems

LLM teammates retain performance after swaps but coordinate less efficiently

Current robot models falter as spatial and procedural complexity rises

Structured agent memories withstand model upgrades better than compressed notes

Training agents to combine graphical interfaces and commands improves efficiency

Streaming video memory works better when internalized as evolving latent tokens

Personalizing agents through cross-session interaction data boosts task success

Identity preservation remains a distinct challenge for generative image models

Causal framework distinguishes deceptive outputs from deceptive mechanisms in language models

Chain-of-thought reasoning traces are legible but not interpretable

Persistent evidence tracking improves language-agent performance on bioinformatics benchmarks

A Vestigial Attention Branch Signals Which Cache Entries to Keep

Evolving terminal environments keeps training tasks challenging as agents improve

LLM judges as measurement instruments fail reliability tests on shared endpoints

Factorized reinforcement learning improves vision-language models’ spatial reasoning

Memory lets traffic-simulation agents build reusable skills without retraining

Image-only pretraining helps build a strong open image generator

Rule-based memory improves question answering across long conversation histories

GRPO Can Reward Lucky Guesses as If They Were Reasoning

Dynamic rubrics improve credit assignment for long-horizon agent training

Graph-based policy constrains LLM agents for topology-aware incident response

Coding agent benchmarks miss review compliance gaps

Digital twins make agent decisions replayable, but hidden tool states remain

Malicious plugin updates can hijack AI agent lifecycle hooks

Post-training choices reshape refusal circuits but leave safety trade-offs

Discrete diffusion accelerates language models while preserving their autoregressive distribution

Coding agents combine generated imagery with editable web-based visual layouts

Distillation Before Reinforcement Learning Improves Reasoning Model Post-Training

Shared infrastructure spread both cheating and resistance through an AI swarm

Agent Security Shifts from Prompt Defenses to Stateful Runtime Control

Graph Machine architecture uses dynamic sparse routing to handle large state efficiently

Deterministic guardrails expose gains that fool LLM-based judges

Agents benefit from representing what expands future planning capacity

Structured behavioral abstractions improve diagnosis of failures in LLM agents

Sparse verifier feedback makes broad credit assignment outperform turn targeting

Adding Depth Helps Video World Models Control Robots More Reliably

Distilled repository skills improve agents conducting machine-learning research

Civilization benchmark exposes agents’ weak monitoring and long-term follow-through

Branching Across Hypotheses Helps Research Agents Avoid Early Search Traps

Reusing agent trajectories cuts token use on repeated tool tasks

LLMs can build agent harnesses, but improvements transfer poorly

Typed trace folding helps agents and observers manage long runs

Layered agent harnesses improve software over repeated autonomous development cycles

Embedding retrieval ranks matching words above shared underlying structure

Language enables precise character and camera control in video worlds

Agents learn selectively from imperfect vision-language model teachers

Long-video agent adapts retrieval to each question’s evidence needs

Coding Agents Need Semantic Measures of Working Memory Performance

Five-level ladder maps reasoning models beyond direct human oversight

Privileged supervision improves action-level credit for language-model agents

LLM agents do not reliably improve from their own experience

Adaptive logit tilting exposes rare language-model behaviors more reliably

Models struggle to improve themselves from vague capability goals

Paper-derived rubrics improve AI generation of scientific research plans

Targeted reflection improves multi-agent systems by locating decisive errors

Models remain weak at steering coding agents through long tasks

Safe agent self-modification requires verifiable, expressive recovery mechanisms

Simulated user personas improve agents’ handling of multi-turn workflows

Tool-call traces expose extraction failures that source fidelity misses

Agents learn to manage long contexts through fine-grained reinforcement learning