Occamy-1.0: Open Pareto-frontier 35B Intelligence for Co-work
Wenhui Chen, Shiwen Cheng, Hao Dong, Chenda Duan, Ruixiang Feng, Zhong Guan, Boqiang Guo, Xueyuan Han, et al.
cs.AI
Quick verdict
Claims 35B open model at low-cost Pareto knee for co-work via SFT+RL+merge; transparent evals, but single-run scores and affiliated benchmarks warrant caution.
Summary
Occamy-1.0 is a 35B-A3B MoE co-work agent model produced by further post-training the released Qwen3.6-35B-A3B checkpoint. The recipe trains two specialists — a Marathon Expert (SFT then HDPO RL) on long-horizon tasks and a Sprint Expert (SFT) on shorter agentic tasks — merges them via a uniform model soup, then applies SAO RL with context compaction. Supporting contributions include execution-grounded synthetic task pipelines with verification/admission gates, token-exact trajectory capture (TiTO), and state replay across three harnesses; the Dressage RL infrastructure, weights, and a subset of SFT data are released. On 12 benchmarks, Occamy leads the 35B-A3B group on Claw-Eval (82.2 vs 69.5 for the base, pass3 71.4), WildClawBench (49.16), and AutomationBench strict pass (27.6 vs 7.5), improving over the starting checkpoint on all supporting skills (e.g., IFEval 91.53, Terminal-Bench 2.1 59.0). Under a frozen pricing protocol it sits at the low-cost knee of the aggregate cost-performance frontier (~$0.058/task vs ~$0.79 for GPT-5.6 Sol). Limitations: most evaluations are single-run without confidence intervals; several co-work benchmarks have organizational links to the team (CommerceAgentBench under Accio-org); the GDPval score is a modified independent reproduction; and only a subset of training data is public.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● low 2
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Staged post-training (Marathon Expert via SFT+HDPO, Sprint Expert via SFT, uniform merge, then SAO RL) turns Qwen3.6-35B-A3B into the strongest comparably sized co-work agent (Claw-Eval 82.2 vs 69.5 base; pass3 71.4 vs 54.8).
- Under the paper's frozen pricing protocol, Occamy-1.0 sits at the low-cost knee of the aggregate cost-performance Pareto frontier (~$0.058/task macro vs ~$0.79 for GPT-5.6 Sol).
- Co-work specialization does not degrade supporting capabilities: AutomationBench strict 27.6 (base 7.5), BFCL v4 65.40 (63.19), VitaBench 41.75 (34.25), TB 2.1 59.0 (49.5), IFEval 91.53 (86.90).
- Token-exact TiTO capture plus state replay makes history-rewritten, multi-harness episodes trainable without corrupting credit assignment across segments.
- Confirmed reward hacking at this scale stems from data/evaluation-infrastructure defects rather than strategic model deception (4 of 4,352 audited trajectories, all from one task).
Strengths
Unusually complete, auditable evaluation protocol
Appendix B pins serving (SGLang, mem-fraction 0.85, shared 262,144-token budget), per-benchmark harnesses, integrity checks for endpoint contamination and trial completeness, and freezes pricing (Table 8) with exact per-task costs (Table 9) behind every Figure 2 point.
Reliability accounting beyond scores
Table 6 decomposes gains: execution success 62.81%→77.55%, 19.5% fewer tokens, 15.2% fewer tool calls, trace wall time 74.58s→39.96s, timeout 9.88%→2.18%, with infrastructure failures separated from model failures.
Serious reward-hacking audit
Repeat-metric discipline on the headline claim
Reproducibility of training itself
Weaknesses
No variance or significance reporting
Organizational proximity to evaluated benchmarks
Pareto-knee framing is protocol-dependent
Mixed-provenance baseline table
Merge/SAO trade-offs disclosed but not ablated
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
Occamy-1.0 (arXiv 2609.11977) is a system report from the 'Accio Team': a 35B-A3B MoE co-work agent model created by further post-training the public Qwen3.6-35B-A3B checkpoint. Three claims anchor the paper: (1) staged post-training — a Marathon Expert (SFT then HDPO RL) for long-horizon tasks and a Sprint Expert (SFT), merged via uniform model soup and refined with SAO RL — makes it the strongest comparably sized model on a 12-benchmark suite (Section 6, Table 5); (2) under a frozen pricing protocol it occupies the low-cost knee of the aggregate cost-performance Pareto frontier (Figure 2, Section 6.6); (3) specialization preserves supporting capabilities (Section 6.4). The infrastructure contribution — token-exact TiTO trajectory capture, state replay, multi-harness adapters, released as Dressage — is arguably as substantial as the model itself.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is Occamy-1.0?
An open-weights 35B-A3B mixture-of-experts model for 'co-work' — multi-step agent work combining tool use, coding, file manipulation, and information gathering. It was created by further post-training the public Qwen3.6-35B-A3B checkpoint, and the weights, part of the SFT data, and the Dressage RL infrastructure are released (HuggingFace Accio-Lab/Occamy-1.0, GitHub Accio-Lab/Dressage).
How does Occamy-1.0 compare with much larger frontier models?
Under the paper's own evaluation protocol it is close to GPT-5.6 Sol, Qwen3.8-Max, DeepSeek V4 Pro, and GLM-5.2 on several co-work benchmarks — e.g., Claw-Eval average 82.2 vs GPT-5.6 Sol's 81.8 — while costing far less per task (~$0.058 vs ~$0.79). It still trails them on knowledge-intensive and simulated-user tasks such as OfficeQA Pro (48.1 vs 74.4) and τ3-Bench Banking (37.1 vs 54.6).
Is the cost-efficiency (Pareto frontier) claim trustworthy?
It holds under the paper's frozen protocol: a September 2, 2026 provider price snapshot, measured token usage, and one shared price vector applied to all six 35B-A3B models. The authors themselves call it a size-matched monetary proxy, not a total-cost-of-ownership claim, and the normalization depends on which ten models are compared.
What training method was used?
A staged recipe: SFT for two specialists (a Marathon Expert on long-horizon tasks additionally trained with HDPO reinforcement learning, and a Sprint Expert on shorter agentic tasks), a uniform parameter-space merge of the two, and a final SAO reinforcement-learning stage with context compaction on broad co-work tasks. Full hyperparameters are published in the paper's appendices.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- The provided text is a sampled excerpt (head and tail kept, marked [...sampling gap...] omissions); middle portions of Section 5.2 (RL math), parts of Section 7 and Appendices A.5–A.6, and a segment of the reference list were not visible, so statements about those parts are unverified.
- Figures 2 and 6–9 were available only as captions and axis labels; numeric placements cited here come from Table 9, which the paper states underlies those figures.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.