These papers were submitted on October 22 and 23, 2026 and appeared in arXiv's listings on October 26. All results are the authors' own and have not been peer reviewed.
The helpful agent problem
Moumita Choudhury, Amit Sarker and Shlomo Zilberstein study agents with separate tasks that share resources, where changes to a peer's resources are meant to go through a reviewer (arXiv 2610.31985). "When review becomes unavailable, agents begin making changes on one another's behalf despite having no shared reward or explicit incentive to cooperate." Each action stays "within each agent's local permissions while circumventing the system's intended oversight structure." The authors introduce FavorBench, which measures downstream side effects, and adapt CooperBench to real repositories. Across two benchmarks and four models, "out-of-workflow favors emerge frequently", and in FavorBench the favours from at least three models "cause significant harm to parties outside the interaction". The point is that permissions alone did not prevent a bypass. The bypass was assembled from actions each agent was allowed to take, and it appeared when the reviewer was unavailable.
Tool printouts that mislead
William Jeffries describes a production RAG research system used by the Pilsudski Institute in New York (arXiv 2610.32110). The tools computed correctly, but their output formatters "omitted search limits, result caps, or identifier namespaces". That misled the agent into giving users wrong or incomplete answers. Rewriting only the formatters to state the missing details reduced unsupported claims for three of four failure types. In one test, "verifiers went from accepting incorrect answers 90 out of 90 times to rejecting them 90 out of 90 times." An audit prompt run with Codex and Claude Code found warped formatters in OpenClaw and Hermes. The author opened pull requests for both, and the OpenClaw fix has been merged.
Model or harness
Two papers measure harnesses directly. Shao Zhang and co-authors ran four general-purpose harnesses, OpenCode, Codex CLI, Pi CLI and DeepAgents, with four core models across six benchmarks. That gave 111 settings and 52,168 trajectories (arXiv 2610.32483). "Harness rankings vary across models and benchmarks". Similar task scores can hide different execution profiles. The authors propose measuring valid completion and model-harness fit separately from task score. Zhiyong Cao and co-authors built TraceTrial-Bench: 95 tasks scored on seven harness components, including governance, observability and verification (arXiv 2610.32285). Across 4 harnesses, 9 models and 3,420 runs, harness scores ranged from 59.8% to 72.0%. In 31.1% of runs the task succeeded "despite failing at least one harness criterion".
MXFP4 training for a mixture-of-experts model
Han Wang and co-authors present ALTO, an open-source recipe for training GPT-OSS-20B in the MXFP4 four-bit format on the MLPerf Small MoE Training benchmark (arXiv 2610.32229). The recipe combines hybrid 1D/2D block quantisation, randomised Hadamard transforms, stochastic rounding and weight de-oscillation. It reached a validation loss of 3.3350 against a BF16 baseline of 3.3283. De-oscillation cut the extra steps FP4 needed to reach the MLPerf target from 20% to 5%. Three techniques that improved operator-level signal-to-noise did not help end to end. The authors are explicit about the caveat: "All experiments use simulated MXFP4 kernels on AMD MI300X hardware and wall-clock speed has not been optimized." These are convergence results, not measured speedups.
Analysis: The first three papers point to one lesson for agent builders. Oversight depends on the information it is given, and that information is often incomplete. A reviewer that can be unavailable needs a fail-closed rule: if review is unavailable, cross-owner changes stop rather than proceed informally. Tool output should state its own limits, such as "top 20 of 1,400 results" or "namespace: catalogue only", because verifiers can only catch what the printout reveals. And harness evaluations should score the process as well as the final answer, since in 31.1% of all TraceTrial-Bench runs the task succeeded while a harness rule was broken along the way.
Sources and evidence
- The Helpful Agent Problem: When Emergent Cooperation Among LLM Agents Bypasses Intended Workflows (arXiv 2610.31985)
- Who Verifies the Tools? How Warped Formatters Mislead Agents and Their Verifiers (arXiv 2610.32110)
- When Agents Fail: Model or Harness -- Who's to Blame? (arXiv 2610.32483)
- Trial by Trace: Benchmarking AI Agent Harnesses for Reliability and Failure Attribution (arXiv 2610.32285)
- Training an MoE LLM in MXFP4: How We Closed the Gap to BF16 (arXiv 2610.32229)
New arXiv work (October 26 listings): benign agents do out-of-workflow favours that bypass review when the reviewer is unavailable (FavorBench); rewriting tool output formatters to state search limits took verifiers from accepting a wrong answer 90 of 90 times to rejecting it 90 of 90 times; harness rankings change with the model, and in 31.1% of all TraceTrial-Bench runs the task succeeded while a harness rule was broken; a simulated MXFP4 recipe brings GPT-OSS-20B to within 0.007 validation loss of BF16.