Wink Pings

Agents start dividing work themselves, but the money goes to reading context

In DAIR.AI’s November 9–15 paper weekly, s3, SoMem, Agency, WikiTrace, and a from-scratch pretraining RSI experiment lay out agent organization, memory, cost, and verification. The numbers are more interesting than the slogans.

In DAIR.AI’s November 9–15 paper weekly, the featured summary lists 7 items and the main text expands to 10. Looking at the ten papers together, agents dividing their own work, organizing their own memory, and checking their own experience became the main thread. Another thread is more practical: cost. A lot of money is spent on reading context and repeated tool calls.

s3: Roles are not written by humans

s3 from the University of Washington and the Allen Institute for AI lets a group of copies of the same model learn how to organize themselves. There is no preset leader, roles, or speaking order; each agent on a task copy decides when to work alone, when to send messages, and when to read messages. All agents receive the same team reward, every token carries the team advantage, and task actions and communication are trained together.

With matched training and inference budgets, evaluated with twice as many agents as during training, it solves coordination problems more than 25 times more often than the strongest baseline. In parallel exploration, agents that communicate do nearly 50% less duplicate work and score up to 9.4 points higher than best-of-N. The experiments use small Qwen3 0.6B and 1.7B models, showing that coordination ability can be learned from a shared objective rather than being fully hard-coded in the harness.

SoMem and WikiTrace: More memory is not necessarily better

SoMem from City University of Hong Kong and Tencent Jarvis Lab lets agents manage memory with ordinary documents. Summaries link back to the original interactions, and original evidence is read-only; at runtime it checks every edit, then reconstructs timelines, relationship graphs, and recent-memory views; when answering, the agent chooses what to search and read, following document links to find evidence. It beats the strongest baseline by 2.66 points on LoCoMo, 3.10 points on BEAM-100K, and 2.00 points on LongMemEval-S.

WikiTrace from KAIST focuses on memory-induced degradation. A compiler turns each past run into success or failure diagnoses and short insights, with insights linked to diagnoses and raw trajectories. New tasks start from all insights; when details are missing, conflicting, or come from failed runs, the agent follows links and uses terminal tools to supplement evidence. On ALFWorld, ScienceWorld, PlanCraft, AppWorld, and τ³-bench Retail, it averages 10.7 points higher than a memoryless agent and 3.2 points higher than using insights alone; the failure rate on originally solvable tasks drops from 6.5% to 4.5%.

Agency: Look at the bill first

Agency from UC Berkeley and UIUC found that, with the same model and task, harness changes can make costs differ by 2.6× while accuracy barely moves. Reading existing context accounts for up to 73% of total cost, even with caching. Tool calls increase cost but not tokens; on average, agents use only 27% of tool output, and in retrieval-heavy tasks up to 89% of turns are spent searching the filesystem. Agent trajectories contain an average of 1.3 to 2.7 independent subtasks that could be parallelized.

Agency sits between existing harnesses and models such as Claude Code and Codex, monitoring and modifying context, tools, and environment. Five conceptual techniques cut cost by up to 35% with little accuracy loss. These techniques are deterministic; they do not require manually resetting sessions or training models to discard context.

RSI: Pushing self-improvement into pretraining

SII and GAIR apply recursive self-improvement to from-scratch pretraining. Pretraining costs hundreds of thousands of GPU hours, lasts months, and cannot be repeated; the real target is the post-training space, which a single score cannot reveal. They break RSI into small loops that can close: agents operate infrastructure and data, small SFT checks the post-training space, expert judgment is converted into data-pipeline metrics, and shared context stores negative results.

The output is 801-30B-A3B-Base, a 30B MoE model trained on 11.8T tokens over three months by a new team. As a base model, it surpasses Qwen3-30B-A3B and Nemotron-3.5-Lightning-30B-A3B and trails only the larger Qwen3.5-35B-A3B; with 6B tokens of SFT and no RL, its average score on four agentic benchmarks is higher than that of a fully post-trained model. The base model and SFT data are open-sourced.

The other papers are no lightweight either

  • Verified Quantum Programs: Google DeepMind and Google Quantum AI build a quantum program library in Lean. A Gemini 4 agent team writes programs and proves them correct; the Lean kernel checks correctness and cost, and a separate comparator performs secondary verification. In the results, parallel search improves by up to 2.55× and quantum arithmetic by 290×; fusion, photochemistry, and drug-metabolism simulations improve by 11×, 8×, and 2.5×, respectively. Proof checking replaces simulation as the reward signal.
  • LongCapHarness: Microsoft Research Asia wraps a harness around frozen models without training them. Llama3-8B gains an average of 64.18 points without training; when the harness is transferred to Qwen2.5-3B, it outperforms NExtLong and PolicyLong—which rely on data synthesis and training—on HELMET and RULER, at less than half their estimated cost. The harness was written by GPT-5.6-Sol; having a small model write it itself causes a significant score drop.
  • ReproBridge: Google turns a successful reproduction into a reusable package, including cleaned-up code, environment configuration, launch scripts, and written guidance. On a benchmark of 1,029 tasks from 36 papers, the refined artifacts raise the reproduction success rate of four models from 76.4% to 92.6% and reduce API spending by 19.7%; an independent auditor checks that the runs actually happened. The gap between the best and worst models shrinks from 21.6 points to 3.9 points.
  • AutoRefine: Meta uses workflow signals such as whether tests pass and whether a change is merged or abandoned to refine the prompt of a code judge; a teacher model organizes labels and can abstain. On Meta’s deployed performance-optimization agent, agreement between the judge and real outcomes rises from 0.16 to 0.51.
  • LingBot-Touch: Robbyant pretrains a tactile foundation model on 4,983 hours of real tactile data, mapping piezoresistive, magnetic, and visual tactile sensors to the same virtual fingertip. For XHand1, whose sensor model was unseen, a frozen model plus a small trained module achieves an average contact-center error of 0.16 mm, while FTP-1 with a dedicated adapter is 5.89 mm. Across six real tasks, LingBot-VLA averages a 57.21% success rate, rising to 75.04% with LingBot-Touch features.

One takeaway

This week’s papers do not offer a universal answer. Their commonality is breaking agent reliability into inspectable engineering layers: how teams organize, how memory links back to evidence, where cost is spent, and how results are verified. The paper numbers apply only to their respective settings. If you want to reproduce them, read the original papers first.

Original: DAIR.AI’s X article https://x.com/i/article/2122023532167081985

Original post: https://x.com/omarsar0/status/2122028163974758735

发布时间: 2026-11-16 09:31