PaperlayerView exact revision on arXiv

Source-linked AI summary

ZGCM-1: A Fully Open and Extremely Efficient Foundation Model for Math and Agentic Search

Jiyan He, Guang Liang, Hao Liu, Haoxiang Guan, Jinbo Sun, Junyi Guo, Wenjun Feng, Yantai Xie, Yifei Shen, Bin Shao, Chuyang Wei, Kai Chen, Kexin Zhou, Minghang Zhu, Shuxin Zheng, Tie-Yan Liu, Taine Zhao, Wenhui Zhu, Xueyin Xu, Xiaoqing Zhang, Yatao Li, Yuxuan Ren

arXiv:2609.13356v1cs.AI

TL;DR

ZGCM-1 addresses the scale and opacity barriers that limit compact researchers’ access to frontier reasoning and agentic-search research. It presents a fully open 7.39B model combining deliberate internal reasoning, active tool use, efficient hybrid-attention training, and open development artifacts. ZGCM-1-7B reports competitive or state-of-the-art results among sub-10B models and contends with much larger systems on challenging reasoning and agentic-search benchmarks, while remaining limited on closed-book recall and broader software agency.

  • Problem

    Frontier reasoning is widely associated with hundred-billion-parameter systems, while competitive models often keep training recipes, curricula, and interaction traces proprietary.

  • Method

    The paper trains a fully open 7.39B dense model that couples deliberate internal reasoning and active external tool use with efficient hybrid attention, FP8 Muon optimization, MDP mid-training, and AI-native R&D.

  • Results

    ZGCM-1-7B achieves competitive or state-of-the-art results among sub-10B models and stands toe-to-toe with much larger systems, including 75.0% on AIME 2026, 63.1% on WebWalkerQA, and 62.0% on Binary Function Search.

  • Takeaways & Limitations

    The results support studying compact models that combine reasoning with external interaction, while the released weights, checkpoints, code, data recipes, logs, and evaluation suites enable open replication.

  • Takeaways & Limitations

    On purely closed-book, recall-intensive tasks without retrieval, ZGCM-1 trails massive frontier systems, and broader repository-level engineering and unstructured terminal navigation remain challenging.

Abstract

from arXiv · show

In this work, we present ZGCM-1, a fully open 7B dense foundation model trained from scratch with extreme data, system, and algorithmic efficiency. ZGCM-1 is founded on a core premise: compact models cannot passively memorize the open web, but can overcome parametric capacity limits by coupling deliberate internal thinking with active external tool use. To support this paradigm across a 256K context, we develop an end-to-end, high-efficiency open training recipe: Architecture & System Co-design: interleaved gated sliding-window and full attention, and a stable FP8 Muon optimizer; Progressive Curriculum & MDP Mid-Training: context scaling across 16K, 64K, and 256K, and the reformulation of interaction traces into Markov Decision Processes. Furthermore, we establish an AI-native R&D workflow where agent swarms autonomously manage cluster operations, data curation, and rapid diagnostic evaluation. Extensive evaluations show that ZGCM-1-7B is competitive across 7B model family on general benchmarks. On several challenging mathematical reasoning and agentic search suites, it remains competitive with frontier models orders of magnitude larger, such as Qwen3-235B-A22B and GLM-5.1. We also show that our pre-training design offers a ~4.2x efficiency improvement in 16K pre-training time-to-loss. Across the full development lifecycle, we distill eight actionable empirical findings-spanning architectural scaling, SFT quality pruning, long-context generalization, and agentic co-training dynamics. To facilitate community research, we open-source model weights from the pre-training, mid-training, and post-training stages, intermediate checkpoints, training code, per-stage data and data recipes, and W&B logs.

1 Introduction

ZGCM-1 addresses scale and opacity barriers with a fully open 7.39B model built around deliberate internal reasoning, active external tool use, and an efficient open training pipeline. It reports strong reasoning and agentic-search performance alongside an AI-native R&D workflow and broad open-source releases.

  • Our Motivation & Core Thesis: The paper frames compact models as limited by static parametric capacity but able to overcome that limit through deliberate internal thinking and active external seeking.
  • Our Motivation & Core Thesis: Rather than passively memorizing the web, ZGCM-1 couples long-horizon reasoning with autonomous web, terminal, and binary-analysis tool use.
  • High-Efficiency Open Recipes: The project develops an efficient open pipeline for 256K contexts using hybrid attention, FP8 Muon optimization, TWEO regularization, and curriculum-based MDP mid-training.
  • Empirical Feasibility: ZGCM-1-7B ranks first on average across 14 reasoning benchmarks at the 7B–8B scale and reaches 75.0% on AIME 2026, 97.1% on MATH-500, and 70.4% on HMMT 2025.
  • Empirical Feasibility: ZGCM-1 contends with much larger frontier models on agentic search, scoring 63.1% on WebWalkerQA, 19.4% on BrowseComp, and 62.0% on Binary Function Search.
  • AI-Native R&D: The work integrates researcher-directed agents across data processing, experimentation, evaluation, deployment, and rapid diagnostic feedback.
  • Empirical Findings and Open Science: It distills eight empirical findings and releases staged model weights, checkpoints, training code, data recipes, logs, and evaluation harnesses for community research.

2 Architecture

ZGCM-1 uses a 32-layer hybrid attention architecture that combines gated 128-token sliding-window attention with periodic global attention. Experiments select the 5:1 schedule because it preserves competitive loss while improving long-context throughput and reducing KV-cache memory.

  • Model Overview: The model is a 7.39B decoder-only Transformer with GQA, RMSNorm, SwiGLU, RoPE, 32 layers, and a maximum context length of 256K tokens.
  • Model Overview: ZGCM-1 interleaves 27 gated SWA layers with five global causal-attention layers in a 5:1 local-to-global pattern across 32 Transformer layers.Global layers are placed at layers 6, 12, 18, 24, and 30, followed by two trailing local layers.
  • Model Overview: Within each gated SWA module, parallel query, key, value, and gate projections produce a locally attended output whose elements are sigmoid-gated before output projection.Global-attention layers omit this gating and attend over the full context.
  • Hybrid-Attention Experiments: The 5:1 SWA schedule achieves 9,566 tokens/s/GPU while matching full attention’s tail loss of 1.93, motivating its adoption for production.SWA 3:1 reaches lower tail loss of 1.92 at slightly lower throughput, while MLA provides 7,645 tokens/s/GPU without corresponding loss improvement.
  • Hybrid-Attention Experiments: At 256K context, SWA 5:1 reaches 3.94× the throughput of full attention, up from 1.13× at 4K.The throughput gap widens as context length increases, making gated SWA suitable for long-context training and inference.
  • Hybrid-Attention Experiments: The hybrid design reduces per-token KV-cache footprint from 128 KiB to 20 KiB by keeping 27 SWA layers local while five global layers retain the full sequence.At 256K context, full attention requires 32.0 GiB and GDN 3:1 requires 7.0 GiB; the supplied passage truncates the hybrid architecture’s final cache value.

3 Pre-Training

ZGCM-1’s pre-training combines broad general-language, knowledge, mathematics, and code learning with progressively longer-context mid-training and efficiency-oriented optimization. Curriculum ordering improves technical-domain BPB at a modest general-domain cost, while FP8/TWEO and the broader co-design improve training efficiency and support diagnostic monitoring.

  • 3 Pre-Training: The two-phase pre-training pipeline builds broad capabilities first, then introduces denser reasoning, instruction, and agentic data while extending context from 16K to 64K and 256K.General Pre-Training uses two stages; Mid-Training retains causal language modeling while scaling context progressively.
  • 3.1.2 Curriculum Pretraining: Lexical-complexity ordering lowers coding BPB from 1.99 to 0.81 and mathematics BPB from 0.97 to 0.94, while general-benchmark BPB rises by 0.03 to 0.09.The ordering applies only to Stage 1; later stages use the full mixture without complexity ordering.
  • 3.1.3 Hyperparameters: TWEO regularization with delayed scaling prevents numerical divergence in FP8 training while sustaining 60% BF16-equivalent MFU at scale.The finding identifies intermediate activation outlier suppression as important for FP8 stability.
  • 3.1.3 Hyperparameters: ∼4.2× speedup in 16K pre-training time-to-loss results from co-designing hybrid SWA, FP8 mixed precision, Muon, and Pre-LN over a BF16/AdamW baseline.The reported component factors are 1.4×, 1.5×, 1.8×, and 1.1×, respectively.
  • 3.3 Monitor During Training: Checkpoint evaluations show heterogeneous capability development: knowledge and QA generally improve, while mathematics and coding gains occur in different intervals with larger short-term fluctuations.The curves are treated as descriptive diagnostics of domain-dependent learning rather than uniform progress measures.
  • 3.3 Monitor During Training: Periodic CUDA allocator cache clearing and garbage collection restored throughput from 554 to 585 model TFLOP/s/GPU after a gradual decline.The fix was deployed in the production training loop.

4 Post-Training

Post-training combines SFT and mixed RL, using curated general and agentic data with quality filtering, balanced reasoning mixtures, and aligned tool protocols. The reported findings emphasize quality over volume, general-data grounding for agents, calibrated long-CoT supervision, cross-modal transfer, and short-SFT activation of long-context capabilities.

  • 4.1.1 General Data: Pruning roughly half of the SFT candidates improves the six-benchmark mean from 67.78 to 68.83, with BBH reasoning gaining +10.10 while some tasks decline.The pipeline uses rule checks, model-based scoring, tiered selection, verification, and decontamination before training.
  • 4.1.1 General Data: Oversaturating long-CoT trajectories degrades general instruction following, while dynamically calibrated mixtures preserve instruction adherence without removing long-CoT reasoning gains.The mixture balances long-form reasoning, direct-response QA, and strict formatting directives.
  • 4.1.1 General Data: Joint think and no-think SFT enables runtime switching between deliberation and concise responses, with reasoning exposure sharpening no-think accuracy across coding, mathematics, and logic.Intermediate checkpoint evaluations report positive cross-modal transfer.
  • 4.1.2 Agentic Data: Agentic SFT aligns schemas with downstream environments across deep research, software engineering, and terminal interaction, supervising tool selection, information gathering, environment interaction, and final responses.Tool-protocol validation checks structured calls, observations, identifiers, termination, and repairability.
  • 4.1.2 Agentic Data: Agent-only fine-tuning degrades interaction fidelity, whereas co-training with general instruction data provides grounding in reasoning and precise constraint following.A joint schedule interleaving general and agentic examples outperforms sequential training and is used for the final run.
  • 4.1.2 Agentic Data: Long-context mid-training allows robust 256K reasoning and agentic behavior to be elicited with moderate-length SFT data, bypassing expensive 256K-long post-training trajectories.The claim is presented as a finding about activating and extrapolating capabilities established during mid-training.

5 Evaluation

ZGCM-1 is evaluated in think mode on broad 7B–8B reasoning and general-capability benchmarks plus web research and binary-analysis tasks. The evaluation emphasizes reproducible harnesses, long-context sampling, iterative tool use, and held-out binary-function recovery.

  • 5.1 Setup: Evaluation runs use a 262,144-token total context budget, with up to 4,096 prompt tokens and 258,048 generated tokens; non-agentic results generally report mean pass@1 over 32 runs.Exceptions include benchmark-specific aggregation rules and specified four-run or single-run means.
  • 5.1 Setup: ZGCM-1 is compared with six reasoning models at 7B–8B scale across 20 benchmarks spanning reasoning, general capabilities, code, knowledge, and instruction following.Reported results use the released 256K SFT checkpoint and think mode.
  • 5.2 Agentic Evaluation: Web-environment deep research evaluates iterative search, inspection, and evidence synthesis against specialized, open-weight, and proprietary systems using WebWalkerQA, BrowseComp, and GAIA.The setup uses a thinking-enabled ReAct policy with at most 64 search-and-read steps.
  • 5.2 Agentic Evaluation: Binary Function Search tests recovery of a target function from a stripped ELF binary using a behavior description and a structured Ghidra-backed interaction protocol.The benchmark contains more than 11,000 validated tasks, while evaluation uses a 50-task subset with projects held out from training.

6 AI-Native Research and Development

ZGCM-1 uses researcher-directed agents and shared human and machine context across the R&D lifecycle. The workflow enables strong autonomy in operational tasks, but design tasks remain more dependent on human direction.

  • AI-Native R&D Workflow: Agents contribute across data, architecture, algorithms, experimentation, infrastructure, evaluation, and deployment, with reusable work and researcher discussions feeding a shared harness.The harness supports agents in executing tasks, inspecting feedback, and iterating.
  • AI-Native R&D Workflow: AI agents can conduct semiautonomous research by searching literature, managing clusters, launching runs, analyzing results, and iterating with limited continuous supervision.During development, agents profiled storage-cluster data transfer and adjusted the training framework’s I/O pipeline.
  • AI-Native R&D Workflow: ACE provides rapid, fine-grained diagnostic feedback that localizes regressions across atomic capabilities during development and ablation studies.Its 2,503 probes cover 183 capabilities across 18 categories, and a complete run takes approximately two to three minutes.
  • AI4AI Autonomy Across the R&D Lifecycle: The autonomy assessment is based on contributor judgments rather than standardized measurements; only one of 99 ratings is L5 and no task exceeds L4.Because assigned levels range from L2 to L4, autonomy is better described per task than as one workflow-wide maturity level.
  • AI4AI Autonomy Across the R&D Lifecycle: Autonomy is task-dependent: experimentation, monitoring, and deployment reach L4, while architecture and learning-algorithm design are L2 and never exceed L3.The assessment covers 11 task categories and uses a five-level rubric, with L4 defined by independent execution within human-defined objectives.

7 Conclusion, Limitations, and Future Directions

The conclusion presents ZGCM-1 as a fully open 7.39B model that combines efficient architecture, training, post-training, and AI-native R&D. It reports competitive results on difficult reasoning and agentic benchmarks while identifying limits in closed-book recall, instruction adherence, broader software agency, and environment robustness.

  • Conclusion: The recipe combines 5:1 gated sliding-window and global attention, FP8 precision, Muon, TWEO, progressive 16K-to-256K context scaling, MDP supervision, and calibrated post-training.The architecture delivers 3.94× throughput and 6.4× KV-cache reduction at 256K, while the pipeline reports approximately 4.2× pre-training time-to-loss acceleration over BF16/AdamW baselines.
  • Conclusion: ZGCM-1-7B achieves competitive or state-of-the-art results among sub-10B models, including 75.0% on AIME 2026, 63.1% on WebWalkerQA, and 62.0% on Binary Function Search.The model is reported to stand toe-to-toe with models orders of magnitude larger.
  • Open Science and Future Directions: The project releases model stages, checkpoints, training code, data recipes, logs, and evaluation suites to support transparent and reproducible research.Future directions include sparse MoE scaling, interactive agentic RL, dynamic retrieval, and more autonomous AI4AI workflows.
  • Limitations: The model’s static memory remains bounded by its 7.39B dense scale, so it trails massive frontier systems on purely closed-book, recall-intensive tasks without retrieval.Deliberate thinking and external search compensate for many factual omissions but do not remove this parametric knowledge bound.
  • Limitations: Heavy reasoning supervision can induce verbosity and slightly impair strict non-reasoning instruction following when calibration is insufficient.The cited examples include IFEval and complex surface-level constraints.
  • Limitations: Broader repository-level software engineering and unstructured Linux terminal navigation remain challenging, while agentic execution depends on strict schemas and stable environment feedback.Extreme observation noise, tool-call deviations, or external search latency can disrupt multi-step rollouts.

8 Contributions

The contributions section identifies the project leadership and distributed-training resources supporting ZGCM-1. Training configurations cover 16K general pre-training and 64K/256K long-context qualification on H100 GPUs.

  • Contributors: Jiyan He is the project lead, with eight named core contributors and thirteen additional contributors listed alphabetically.The contributor lists distinguish core contributors from other contributors.
  • Training Configurations: Distributed-training configurations cover general pre-training at 16K context and long-context qualification at 64K and 256K using H100 GPUs with FP8 computation.The configuration terminology includes tensor, pipeline, context, and data parallelism, plus micro- and global-batch sizes.
  • Training Configurations: The staged route to 256K reaches a final loss of 1.16 versus 1.19 for the direct route at comparable throughput.The staged route trains 10B tokens at 64K followed by 20B tokens at 256K, while the direct route trains 30B tokens at 256K.

A.2 Capability Dynamics during General Pre-Training

Capability dynamics are uneven across pre-training: all benchmarks improve, but domain-specific gains continue later than broad gains. Curriculum ordering improves coding and mathematics fit while slightly worsening general-domain BPB, and lexical complexity is unsuitable as a difficulty proxy for technical content.

  • Evaluation Scope: The direct checkpoint evaluation uses fixed subsets, so small 50-item benchmarks such as GSM-Symbolic, Minerva Math, and OpenBookQA have noisy point-to-point changes.Only the overall trend is informative for those benchmarks.
  • General Pre-Training: Every benchmark improves from the first to the last checkpoint window, but improvement rates diverge and most broad-benchmark gains arrive before 2.5T tokens.HumanEval adds 7.08 points and GSM-Symbolic 14.80 points over the final 1.7T tokens, although GSM-Symbolic uses only 50 items and has wide uncertainty.
  • Mid-Training Capability: Short-SFT probes show large downstream-capability gains from the end of General Pre-Training to the end of Mid-Training, including MATH-500 rising from 35.83% to 74.12%.HumanEval+, IFEval Strict, and MMLU also rise substantially, but individual benchmarks are not monotonic at every checkpoint.
  • Mid-Training Capability: The final Mid-Training probe uses three SFT epochs instead of two, making its trajectories diagnostic evidence of downstream potential rather than a strictly matched causal ablation.This probe is distinct from direct base-checkpoint evaluation.
  • Curriculum Pre-Training: Lexical complexity is cheap for ordering general-language data but poorly reflects reasoning difficulty in code and mathematics, where surface statistics can be dominated by boilerplate or token variation.The schedule therefore orders general-language documents while interleaving code and mathematics independently.

B.1 General SFT Data Selection

General SFT selection combines broad capability coverage with structural validation, quality filtering, decontamination, and mixture calibration. Controlled comparisons show that quality-focused pruning and balanced reasoning proportions improve aggregate performance while avoiding instruction-following degradation.

  • General SFT Data Selection: SFT data is organized into general and agentic branches covering instruction following, knowledge, code, mathematics, science, dialogue, reasoning, and tool interaction.Examples are normalized to a common message and tool schema before structural, correctness, and tool-turn filtering.
  • Capability Balancing and Reasoning Mixture: Excessive long-reasoning shares introduce verbosity bias and impair instruction adherence, so the final mixture reduces long-reasoning proportions.The selected composition retains reasoning supervision while balancing long-reasoning, direct-response, and instruction-following examples.
  • Evidence for Quality-First Selection: High-quality SFT data raises aggregate performance despite using fewer examples.The quality-focused version uses 44.9% fewer examples than the minimally processed version and 37.5% fewer than the broadly filtered version, while its six-benchmark mean rises to 68.83 from 67.78 and 67.98.
  • Agentic SFT Data: The agentic corpus includes deep-research, software-engineering, and terminal-interaction trajectories filtered for valid tool use, safety, completion, and trace quality.Execution-grounded and execution-free pools undergo separate structural and quality checks before projection into a shared structured-tool format.
  • Thinking-to-Direct Transfer: Reasoning supervision improves several mathematics and coding evaluations but transfers unevenly to direct-response tasks.AIME 2025 rises from 3/30 to 13/30, AIME 2024 from 2/30 to 10/30, and MBPP+ improves by 20.90 points, while MMLU decreases by 2.60 points and HumanEval+ by 7.32 points.
  • Thinking-to-Direct Transfer: The checkpoint comparison is observational rather than a controlled data ablation, so differences may reflect tokens, composition, history, or optimization schedule.This limits attribution of the reported gains specifically to the compared data mixture.

C Agentic Evaluation Protocol

The agentic evaluation protocol tests deep research and binary-function search through structured, iterative tool interaction. Its datasets, scoring rules, and harnesses enforce exact outputs and reproducible validation against hidden oracles.

  • Protocol Scope: The protocol evaluates agents on deep research and binary-function search with task-specific structured environments and exact success criteria.Deep research uses search-and-read interaction, while binary-function search requires the exact target function-entry ELF virtual address.
  • Deep Research: Deep research permits at most 64 search-and-read steps and uses Serper, Jina Search, Jina Reader, no-thinking summarization, and Qwen3-30B-A3B-Instruct-2507 judging.If the step budget is reached, the harness requests a final answer from the evidence collected so far.
  • Evaluation Caveats: Interpretation of deep-research results depends on search date, provider behavior, extraction, summarization, judge settings, and retry policy.These protocol factors constrain direct comparisons across runs or environments.
  • Binary Function Search: Binary-function search requires distinguishing a target function in a stripped ELF binary by combining behavioral, control-flow, call, constant, error, and output evidence.The agent uses a Ghidra-backed interface for function enumeration, metadata, decompilation, search, references, and assembly listings.
  • Binary Function Search: The Ghidra workflow iteratively narrows candidates, refines evidence, submits an entry address, and validates it against a hidden oracle and function-boundary checks.Each turn produces at most one structured tool call, creating a reasoning-action-observation loop until submission or the turn limit.

D.4 Evaluation Results

ACE decomposes evaluation into 183 atomic capabilities across 18 categories and compares ZGCM-7B with reference models. ZGCM-7B is strongest in mathematics and several reasoning categories, with its advantage over Qwen3-8B-256K concentrated in selected capabilities.

  • Category-Level Results: ZGCM-7B performs best in mathematics, planning, logic, causal reasoning, and reading, scoring 98.18%, 97.78%, 96.08%, 96.00%, and 95.04%, respectively.Its weakest categories are instruction following at 49.74%, truthfulness at 68.91%, code at 72.24%, and language at 79.75%.
  • Category-Level Results: ZGCM-7B’s gap to the strongest reference models is concentrated in a small set of categories rather than spread across the capability space.ACE therefore highlights capability-specific strengths and weaknesses rather than only a single aggregate score.
  • Comparison with Similar-Scale Model: Compared with Qwen3-8B-256K, ZGCM-7B leads by 31.11 points in long-context processing, 20.74 in arithmetic, 14.45 in planning, and 10.00 in world modeling.The comparison covers the 18 ACE capability categories shown across model rows and category columns.
  • Scope and Limitations: ACE is a diagnostic complement rather than a comprehensive measure of end-to-end application performance.Its results depend on probe coverage and scoring quality, and rubric-based LLM judging may retain residual bias.

E AI Autonomy Rubric and Contributor Assessments

The rubric evaluates AI autonomy across 11 R&D task categories using five responsibility levels, while contributor assessments record judgments rather than standardized capability measurements. It distinguishes autonomy from productivity or quality and cautions against generalizing narrow-task demonstrations.

  • Assessment Scope: The appendix covers 11 task categories across seven R&D domains, including data, infrastructure, architecture, algorithms, experimentation, evaluation, and deployment.The domains include data engineering, infrastructure engineering, model architecture design, learning algorithm design, experimentation and monitoring, model evaluation, and deployment engineering.
  • Assessment Scope: Nine contributors supplied 99 ratings, but these assessments represent contributor judgments rather than standardized capability measurements.Contributor identifiers remain consistent across tasks in the reporting table.
  • Assessment Scope: The five-level rubric progresses from assistance with individual steps to autonomous objective formation and sustained coordination across R&D stages.L2 covers predefined workflows, L3 adaptive execution with approval at consequential decisions, and L4 autonomous iteration within human-defined objectives and constraints.
  • Assessment Scope: Autonomy levels do not measure productivity, output quality, or scientific novelty, and higher autonomy does not necessarily imply better results.Lower-level assistance can still provide substantial benefits, while L2 behavior may also be conventional automation.
  • Assessment Scope: L5 is a reference for full autonomy rather than a demonstrated capability, retaining dependence on prior knowledge, operating constraints, and external validation.Its defining distinction from L4 is autonomous objective formation and sustained coordination across stages.
  • Data Engineering: The rubric describes autonomy separately for data cleaning, acquisition, and synthetic-data generation, with higher levels adapting strategies and using downstream outcomes to guide iteration.These task-specific levels range from human-defined assistance to evolving workflows coordinated with other R&D stages.
  • Model and Algorithm Design: For architecture and learning-algorithm design, higher autonomy involves proposing, implementing, evaluating, and refining variants, while consequential structural or training changes require human approval at L3.L4 covers iterative autonomous search and refinement under objectives and constraints; L5 includes identifying research directions or questions.
  • Infrastructure Engineering: The operator-design rubric extends the same progression to hardware-aware implementation, compilation, tuning, and repair, with consequential hardware-specific changes requiring human approval.Training and inference framework design begins with AI assistance for configuration, integration code, and error interpretation.

F Evaluation Reproducibility and Provenance

Reliable benchmark claims require provenance across the model, data, invocation, and scoring layers, while the reported evaluation uses deterministic no-think decoding. Table 16 presents selected no-think results for ZGCM-1-7B and comparable 7B–8B models.

  • Provenance: Reproducible benchmark results require binding the model artifact, evaluation data, invocation contract, and scoring implementation.The evaluation record should preserve details such as model hash, immutable dataset revision, prompts, decoding controls, extraction, metric implementation, seed, engine, and hardware.
  • Evaluation Protocol: ZGCM-1 is evaluated in no-think mode with deterministic decoding using temperature 0 and top-p = 1.0.
  • Benchmark Results: Table 16 reports selected no-think benchmark results (%) for ZGCM-1-7B and selected 7B–8B models, marking best and second-best results by cell color.
Loading 2609.13356v1…