cs.AIAug 24, 2026
SaveApodex 1.1: Scaling Agentic Intelligence for Complex Work
Abstract
General-purpose language models can reason and synthesize knowledge, but complex work also requires sustained interaction with files, information sources, and executable code, together with state maintenance, failure recovery, and verifiable delivery. We call this working capability: sustained, verifiable progress toward a real-world objective. Apodex 1.1 develops this capability along two complementary dimensions. Environment Scaling expands the diversity and verifiability of executable file, search, and code environments, while Agentic Coordination Scaling trains agents to decompose long-horizon tasks, delegate parallel work, integrate asynchronous results, and replan. A shared execution harness and AgentOS maintain task state and provenance across tools and agents, and training turns environment trajectories and coordination traces into reliable behavior. Across complex professional work, finance, scientific research, mathematics, coding, and search, Apodex 1.1 reaches the leading performance band despite using a substantially smaller model than many frontier systems. The 35B-parameter Apodex 1.1 Mini further retains strong working capability in a locally deployable form. These results ground agentic intelligence in useful, verifiable work completed over time and advance our goal of building a Heavy-Duty Solver for ambitious, long-running tasks.
Figures & tables
Figure 2: The controlled capability-development loop. Environment Scaling and Agentic Coordination Scaling create the executable trajectories used by training; evaluation, real failures, and human feedback determine which capability gaps the next Task Pipeline should target.
| Family | Construction | Verification boundary | Role |
| File worlds | Profession-conditioned multi-format workspaces | Code-derived values or recorded provenance | Authority and transformation |
| Search worlds | Indexed or open evidence sources | Provenance + claim review | Discovery and evidence alignment |
| Code worlds | Repositories and stateful sandboxes | Tests + artifact checks | Executable transformation |
Table 1: Environment families and their assurance boundaries. “Exact” means code-derived or independently reconciled , not accepted solely from a language-model judge.
| Concern | Failure mode | Mechanism |
| Workspace state | Ambiguous ownership or backend-dependent visibility | Stable three-region namespace with explicit private/shared workspace topology |
| Reference library | Durable user documents exposed to mutation | Read-only /shares mount, configured by the deployment |
| Coordination state | Plan and completion state disappear during compaction | External task board, periodic re-injection, and finalization gate |
| Live intervention | A follow-up waits behind a long fan-in operation or races a report | Message-addressable queue, interruptible waits, delivery acknowledgement, and lease renewal |
| Context pressure | Estimate-driven compaction fires too early or too late | Provider-reported trigger, tool-result eviction, and conditional LLM summary |
| Budget enforcement | Hard cancellation discards work in flight | Shared soft/hard budget, deadline-clamped waits, and bounded report recovery |
Table 2: AgentOS 1.1 mechanisms and the failure modes they address.
| (a) Professional Work | ||
| Models | APEX-Agents | GDPVal |
| DeepSeek-V4-Pro ( DeepSeek-AI, 2026 ) | 24.3 | 71.2 |
| Gemini-3.1-Pro ( Google, 2026a ) | 32.0 | – |
| Claude-Opus-4.6 ( Anthropic, 2026d ) | 33.0 | – |
| GPT-5.4 ( OpenAI, 2026c ) | 33.3 | – |
| DeepSeek-V4-Flash-0731 ( DeepSeek-AI, 2026 ) | 34.4 | 72.7 |
Table 3: Performance on professional work, finance, and business simulation benchmarks. Bold indicates the best reported result in each benchmark column. Apodex results are grouped below the horizontal rule. All GDPVal entries report win rate; all external-model GDPVal results shown here are our reproductions under the Apodex harness. The Claude Opus 5 APEX-Agents result is likewise reproduced internally. YC-Bench ( He et al., 2025 ) reports mean final net worth in USD; external YC-Bench results are taken from its official leaderboard. The DeepSeek V4 entries on FrontierFinance are internally reproduced under the stated benchmark metric; the remaining external FrontierFinance entries are reported reference results. † YC-Bench uses the benchmark’s default agent configuration; no separate Agent Team result is reported.
| (a) Scientific Research | ||
| Models | FrontierScience Research | BioMysteryBench Human-difficult Set |
| Gemini-3.1-Pro ( Google, 2026a ) | 16.7 | – |
| Claude-Opus-4.5 ( Anthropic, 2025a ) | 17.5 | – |
| Claude-Opus-4.7 ( Anthropic, 2026e ) | 20.0 | 27.0 |
| GPT-5.2 ( OpenAI, 2026b ) | 25.2 | – |
| Seed2.1-Turbo ( ByteDance Seed Team, 2026 ) | 33.3 | – |
Table 4: Performance on scientific-research, general-reasoning, and deep-search benchmarks. Bold indicates the best reported result in each benchmark column. Apodex results are grouped below the horizontal rule. DeepSearchQA reports F1, and BioMysteryBench uses the Human-difficult subset.
| Models | FrontierFinance | FrontierScience Research | APEX-Agents |
| Proprietary models | |||
| Gemini-3.1-Pro ( Google, 2026a ) | 30.5 | 16.7 | 32.0 |
| GPT-5.6-Sol ( OpenAI, 2026a ) | 46.8 | – | 39.9 |
| Qwen3.7-Plus ( Qwen Team, 2026 ) | – | – | 22.4 |
| Muse-Spark-1.0 ( Meta AI, 2026 ) | – | 38.0 | – |
| Open-weight models | |||
Table 5: Model-scale-efficient agentic performance of Apodex mini models (35B parameters). Models are grouped by availability; Apodex results appear below the final rule. Dashes denote unavailable matched results. Bold indicates the best result in each benchmark column among the selected comparison rows.
| Models | IMO 2025 | IMO 2026 | USAMO 2026 | ProofBench Basic | ProofBench Advanced |
| Reference threshold | 35 | 29 | 25 | – | – |
| Apodex 1.0 w/ Agent Team | 12.5 | 13.0 | 5.8 | 63.3 | 20.0 |
| Apodex 1.1 w/ ReAct | 24.3 | 18.5 | 16.0 | 80.0 | 46.4 |
| Apodex 1.1 w/ Agent Team | 36.5 | 30.5 | 26.5 | 96.7 | 63.3 |
Table 6: Mathematical reasoning results on competition-level proof tasks. IMO 2025, IMO 2026, and USAMO 2026 are scored under the MathArena protocol.
| Benchmark | Model | Score |
| Terminal-Bench 2.1 | Gemini 3.6 Flash | 91.9 |
| Kimi K3 | 88.3 | |
| DeepSeek V4 Flash 0731 | 82.7 | |
| Claude Opus 5 | 77.3 | |
| Apodex 1.1 | 70.8 | |
| SWE-bench Verified | Claude Opus 5 | 92.2 |
Table 7: Coding-agent results on Terminal-Bench 2.1 and SWE-bench Verified. One result is retained per model on each benchmark; where multiple settings are available, the highest reported score is shown.
| Model | Positive (%) | Zero (%) | Negative (%) | Avg. score |
| DeepSeek-V4-Flash-0731 ( DeepSeek-AI, 2026 ) | 75.6 | 19.5 | 4.9 | 54.9 |
| Kimi-K3 ( Team et al., 2026 ) | 75.6 | 22.0 | 2.4 | 60.1 |
| DeepSeek-V4-Pro ( DeepSeek-AI, 2026 ) | 85.4 | 14.6 | 0.0 | 61.3 |
| Claude-Opus-5 ( Anthropic, 2026g ) | 85.4 | 12.2 | 2.4 | 64.4 |
| GPT-5.6-Sol ( OpenAI, 2026a ) | 85.4 | 14.6 | 0.0 | 67.4 |
| Apodex 1.1 w/ ReAct | 75.6 | 19.5 | 4.9 | 57.0 |
Table 8: FrontierSearchBench results: mean normalized task score over 41 tasks on a 100-point scale, with the proportion of tasks scored positive, zero, and negative. Negative aggregate scores are possible. Column best values are in bold.
| Model | Harness | Pass Rate (%) |
| GPT-5.6-Sol | Codex | 20.6 |
| Grok-4.6 | Claude Code | 20.6 |
| Kimi-K3 | Claude Code | 17.5 |
| Claude-Opus-5 | Claude Code | 17.5 |
| Qwen3.8-Max | Claude Code | 15.5 |
| DeepSeek-V4-Flash-0731 | Claude Code | 12.4 |
Table 9: FrontierResearchBench results. Pass Rate is the fraction of tasks awarded full score. Column best values are in bold.
Figure 5: The HDS6 capability taxonomy and process-grading pipeline. The six capability groups contain four rubric items each; an integrity gate is applied independently of the weighted rubric.
Figure 6: HDS6 comparisons between Apodex v1.0 and v1.1 across three evaluation settings. In Deep Discover, v1.0 uses the 397B Agent Team with eight-run aggregation, whereas v1.1 uses a single run. C1–C6 follow the process-capability taxonomy in Figure 5 ; comparisons are intended within, rather than across, panels.
Appendix figures & tables1 asset
Supplementary material from the paper’s appendix.
Appendix
Table 10: Twenty required deliverables plus three auxiliary files. Every TSV row count was read back from the published file.
More in this topic
Long-Horizon LLM Agents
Long-horizon large language model agents plan, use tools, and preserve task state across extended, multi-step interactions.
See all 448 papers