AI research wiki
agentic-systems · 2026-11-06 · Tier 2

Agent shorts: restart monitors, turn budgets, reasoning compaction, latent thinking, and new research benchmarks

Agent shorts: restart monitors, turn budgets, reasoning compaction, latent thinking, and new research benchmarks

Sources: Monitor Accuracy Isn't the Answer, arXiv:2610.36544 · What Does Multi-Turn RL Really Improve?, arXiv:2611.06483 · Why Pay for Long Reasoning, arXiv:2611.03852 · TAC-CoT, arXiv:2611.01926 · Hybrid Computer Use, arXiv:2610.41432 · SleuthVLM, arXiv:2611.06083 · FrontierPhysics, arXiv:2611.06910 · HarnessGrad · raw: raw/huggingface/

TL;DR. The shared theme is measuring what an agent spends against what it actually gets. Restart monitors are judged by the wrong metric. Multi-turn RL gains vanish when turn budgets tighten. Rewriting a model's earlier reasoning into compact notes shortens its later reasoning without retraining. Three new research benchmarks show frontier agents still mostly fail end-to-end research.

Items

  • Restart monitors (2610.36544). A monitor with a fixed 90% accuracy can swing from a 5-point loss to a 35-point gain, depending only on how likely a fresh attempt is to succeed with the remaining budget. In 194 executions on SWE-bench Verified SymPy issues, a 2.29-point apparent gain traced to one surviving attempt, a second model lost 42.29 points, and none of 103 live comparisons met the certification bar. The paper offers a decision-sufficiency test for whether a compressed summary keeps the information needed to act.
  • Multi-turn RL (2611.06483). Pass@(K,T) evaluates sampling breadth and turn depth together. RL's gains depend on sufficient turn budgets and concentrate in error recovery, not better single-step decisions. BudgetMix RL varies budgets in training for robustness.
  • Reasoning compaction (2611.03852). Rewriting prior reasoning into compact state records before tool results arrive usually shortens later reasoning in open-weight models, with less repeated verification. It can cut total spend, though quality and rewrite-inclusive time depend on the target and compressor.
  • TAC-CoT (2611.01926). Replaces explicit reasoning with latent segments and keeps text only for tool calls, observations and answers. Up to +8.8 points and up to 6.8x fewer forward passes.
  • Hybrid Computer Use (2610.41432). Unifies code, accessibility and GUI channels, with batched actions and accessibility diffs to reduce round trips. Top Partial and Binary scores on OSWorld 2.0.
  • SleuthVLM (2611.06083). A vision agent over transcribed speech beats omni models that hear audio: 75.0% on four benchmarks, 2.6 points above Gemini 3.8 Flash. "Hearing is not the bottleneck, agentic ability is."
  • FrontierPhysics (2611.06910). 56 tasks from 42 physicists. No configuration passes more than 46 of 168 rollouts. Weakness is rigor and expert intuition, not knowledge.
  • Science's Last Exam (@omarsar0, @rohanpaul_ai). 113 tasks in 9 disciplines, graded by hidden checks and against studies published after training cutoff. DeepSeek V4.1 Flash led hypothesis generation at 21.28%.
  • DeepFinance (@ycombinator). A long-horizon benchmark of 21 high-finance problems.
  • HarnessGrad (ZJU). A platform crossing 8 harnesses, 19 self-improvement methods and 56 benchmarks, with the model held fixed so only the harness learns.
  • Claude Code status-line fix (@b1rkh0ff). Background subagents were re-reading the entire conversation every 30 seconds to update a status line. The fix cuts that cost about 99%, or roughly 5% of total limit use.

How this relates to prior wiki pages

  • Extends NVIDIA's frequent compaction (11-05) from tool history to the model's own reasoning. The restart-monitor paper's decision-sufficiency test is the missing way to check whether a compaction lost anything that matters.
  • Parallels DEPTH (11-05), which found that task depth, not turn count, predicts difficulty.
  • Confirms agent benchmarks: SemiAnalysis's GDPx (11-05, no agent above 12%) and today's three benchmarks all keep research-grade work far from solved.
  • Tension: on the harness thread, HarnessGrad treats the harness as a learner while KernelSeek (summary) strips it down.

Gaps

The restart study is on 19 SymPy issues only. Reasoning compaction has no fixed quality guarantee. Benchmark scores are initial and self-run by their authors.