Source-linked AI summary
Palmyra x6 Technical Report: An Agentic, Tool-Use Model Post-Trained via Anchored Supervised Fine-Tuning
Peng Du, Kiran Kamble, Rakshith Vasudev, Zhizhuo Yang, Rohith Nadimpally, Arjun Krishna, Waseem Alshikh, Daniel M. Bikel
TL;DR
Enterprise agentic workloads demand strong multi-step planning, tool calling, and long-horizon task completion. This paper post-trains a Mixture-of-Experts model with anchored supervised fine-tuning on verified synthetic tool-use trajectories, achieving gains over the prior default and leading benchmark performance.
Problem
Enterprise agentic workloads require capable multi-step planning, tool calling, and long-horizon task completion in tool-rich environments.
Method
The model uses anchored supervised fine-tuning on verified synthetic tool-use trajectories, with demonstrations, rollout filtering, and trajectory-level verification.
Results
Palmyra x6 leads the prior default across all six benchmarks and tops BFCL Core at 0.785 against five frontier models.
Takeaways & Limitations
The results support a conservative anchored post-training recipe for producing an agentic model competitive with current models.
Takeaways & Limitations
Palmyra x6 is specialized for agentic and tool-use workloads, while non-agentic behavior largely follows the inherited base model.
Abstract
from arXiv · showhide
Palmyra x6 is a large language model optimized for use with enterprise-oriented agentic tasks. The model was built by post-training a Mixture-of-Experts base model with Anchored Supervised Fine-Tuning on a compact corpus of verified, synthetic tool-use trajectories, optimized with a Muon + Adam hybrid. The recipe is deliberately conservative and deliberately controlled: 626 trajectories, a single epoch, a low learning rate, and a KL anchor to the frozen base. The model shows substantial gains over the previous default model for Writer Agent, and compares favorably with several recent models on public benchmarks, scoring the highest on BFCL Core at $0.785$ and posts the highest six-benchmark mean of the cohort. Furthermore, the model has shown itself to be competitive or leading relative to comparators in our bias and safety evaluations.
1 Introduction
The introduction argues that carefully designed post-training can adapt an existing high-performing LLM when paired with high-quality data and a matched objective. It presents Palmyra x6 as a substantial improvement for Writer Agent while emphasizing enterprise-oriented internal evaluation over public benchmarks alone.
- Motivation: Careful post-training of an existing high-performing LLM can suffice when training data quality and objective design match the target needs.This approach can avoid rebuilding the entire pipeline from pretraining through post-training.
- Results: Palmyra x6 delivers a huge improvement over the previous default model for Writer Agent.The improvement is shown in Figure 1.
- Evaluation: Public benchmark performance serves as a sanity check and guide, not the ultimate evaluation rubric.The stated focus is on internal datasets designed to mimic enterprise customers’ domains and use cases.
- Evaluation: The introduction also situates Palmyra x6 against other frontier models available at the time of development.Those comparisons are collected in Figure 2.
2 Overview
Palmyra x6 was developed in two controlled phases, first selecting an ASFT recipe with a KL anchor and then changing only the base model and optimizer for final training.
- Recipe development: Twelve SFT and ASFT runs established ASFT with KL weight K=0.1 over the 12-dataset primary mix as the primary candidate recipe.The matrix varied objectives, KL-anchor strengths, learning rates, epoch counts, and six data recipes.
- Final training and optimization: Palmyra x6 retained the selected data and ASFT objective while changing exactly two elements: the base model and optimizer.The final optimizer used Muon instead of Adam for the 2-D weight matrices.
- Writer Agent evaluation: Palmyra x6 led the previous default Writer Agent model on all six evaluation benchmarks, with largest gains of +0.320 on MCP-Atlas, +0.305 on FinanceBench, and +0.304 on IFBench.Bars show mean score, with higher being better; Palmyra x6 error bars denote one standard error.
- Controlled comparison: The controlled design attributes capability differences between development candidates and Palmyra x6 to the new base model and optimizer rather than data or loss changes.The report states that variable isolation was maintained with one data recipe and one objective.
3 Model Architecture
Palmyra x6 uses GLM-5.2 as its base and preserves the GlmMoeDsa architecture during fine-tuning. Its design combines sparse Mixture-of-Experts feed-forward layers, Multi-head Latent Attention, and a DeepSeek Sparse Attention indexer, with DSA IndexShare enabling cross-layer reuse of selected indexer outputs.
- Base architecture: Fine-tuning updates the GLM-5.2-derived model’s weights without changing its inherited GlmMoeDsa architecture.The architecture is inherited unchanged from GLM-5.2.
- Core components: The architecture combines a sparsely activated Mixture-of-Experts feed-forward stack, Multi-head Latent Attention, and a DeepSeek Sparse Attention indexer.These are the three ingredients identified in the report’s informal architectural summary.
- Attention mechanism: DSA IndexShare reuses the sparse-attention indexer’s selected information across layers.The passage identifies cross-layer reuse of the sparse-attention indexer’s selected output as a significant base-model feature.
4 Training Data
Palmyra x6 trains on a compact corpus of 626 curated, synthetic tool-use trajectories spanning six realistic multi-tool task categories. Data are generated and filtered through demonstration-conditioned rollouts, verification, and leakage controls before training.
- Corpus composition: Every assistant turn is machine-generated, while prompts come from internal Writer subject matter experts and some external benchmark and task suites.Synthetic data include plans, tool calls, and final answers; the task prompts are separately sourced.
- Task coverage: Six task categories exercise realistic multi-tool workflows, including financial research, data analysis and coding, clinical/medical tasks, MCP suites, simulated worlds, and retrieval-augmented generation.The categories were selected to cover varied agentic workflows and tool-use settings.
- Data generation and filtering: Demonstration-conditioned student rollouts are retained only after verification against teacher answers and filtering for demonstration-based cheating.Verification judges the full trajectory, while leakage checks reject explicit demonstration use, high reasoning overlap, or tool-free answers when tools were available.
- Data generation and filtering: The pipeline applies automated quality gates, including required tool use, message limits, encoding hygiene, and independent two-model panel consensus.A stricter strong-panel-majority ablation retained 240 of 626 trajectories.
- Corpus composition: 626 curated trajectories form the primary Palmyra x6 training mix across 12 private datasets.The datasets cover financial research, data analysis and coding, medical-agent tasks, retrieval-augmented generation, simulated worlds, and MCP-related suites.
5 Training Methodology
Palmyra x6 uses Anchored Supervised Fine-Tuning to adapt a 744B-parameter model on 626 trajectories while limiting distributional drift. Its training recipe combines DFT weighting and a KL anchor with a Muon–Adam optimizer split and conservative final settings.
- Anchored Supervised Fine-Tuning: ASFT adapts a 744B-parameter model on 626 trajectories by combining DFT token-probability weighting with a KL anchor to a frozen base.DFT emphasizes tokens the model already predicts confidently, while the per-token k3 KL penalty controls distributional drift.
- Hyperparameter Selection: K=0.1 was the headline KL-anchor setting, paired with the lower learning rate rather than the weaker K=0.02 alternative.The run uses a frozen reference copy of the base model for the KL term.
- Muon–Adam Optimization: Muon updates 2-D attention and FFN matrices with Newton–Schulz orthogonalization and RMS matching, while Adam remains for 1-D and non-matrix parameters.Muon is split across per-head MLA projection sub-matrices rather than treating concatenated heads as single matrices.
- Memory and State Management: Muon uses a single momentum coefficient of approximately 0.95, retains weight decay 0.1, and CPU-offloads optimizer state for the 744B-parameter MoE.The Muon groups omit a second-moment estimate, roughly halving matrix-parameter optimizer state relative to Adam.
- Final Training Configuration: The converged configuration uses learning rate 5 × 10−7, cosine decay to 5 × 10−8, a single epoch, batch size 16, and sequence length 65,536.It uses context parallelism 2 and the primary 12-dataset reasoning-trace mix containing 626 trajectories.
6 Evaluation
Palmyra x6 leads the prior default across all six evaluation benchmarks, with especially large Writer Agent gains, and compares favorably with five widely deployed frontier models. It achieves the cohort’s highest BFCL Core score and six-benchmark mean.
- Prior-default comparison: Palmyra x6 leads the prior default on all six benchmarks, evaluated across finance-agentic, tool-use, and instruction-following suites.Scores use a 0–1 scale; ± values denote reported standard errors, and prior-default figures are means over available runs.
- Frontier-model comparison: 0.785 on BFCL Core is Palmyra x6’s highest score among five widely deployed frontier-model comparators.Figure 2 compares the models across all six benchmarks.
- Frontier-model comparison: 0.765 is Palmyra x6’s highest six-benchmark mean among the frontier-model cohort.The mean summarizes performance across the six benchmarks shown in Figure 2.
7 Security and Safety
Palmyra x6’s safety profile combines conservative training controls and deployment-level guardrails with strong political-bias, refusal-consistency, and adversarial-safety results. The paper nevertheless emphasizes that tool-use risks depend substantially on the deployment environment and that the KL anchor only limits behavioral drift rather than preserving alignment wholesale.
- Training and behavioral drift: The KL anchor constrains Palmyra x6 toward the frozen GLM-5.2 base, limiting behavioral drift without guaranteeing wholesale preservation of alignment or refusal behavior.The authors make the weaker claim that the anchor preserves general capabilities and can reasonably be expected to limit alignment and refusal drift.
- Deployment risks: Agentic risk is concentrated in the tools and untrusted content encountered during deployment, so prompt-injection and misuse defenses must also operate at the harness level.Training domains include web search, code execution, maps and weather, calculator and reference tools, and document retrieval; the Writer platform supplies additional guardrails.
- Bias and refusal consistency: 80% of Washington Post political-bias answers presented both sides, while Palmyra x6 showed the smallest net leftward lean among tested models.The evaluation used 90 answers across 29 questions and three samples per question, with Palmyra x6 in its deployed system-prompt configuration.
- Bias and refusal consistency: 1.2% was Palmyra x6’s political refusal-mismatch rate, the study’s second-lowest, alongside 88% toxic-content safety, control-level over-refusal, and no detected U.S. political or demographic bias.The evaluation compared eight chat endpoints using 19,674 responses and 33,862 blinded judge scorings.
- Adversarial safety: 8.6 points was the FORTRESS adversarial-safety gain from the deployment system message over GLM-5.2, rising from 58.4 to 67.0 while benign performance was 96.4.FORTRESS covered 500 expert-written adversarial prompts spanning CBRNE, political violence and terrorism, and criminal and financial domains, each paired with a benign twin.
8 Limitations •
Palmyra x6 is specialized for agentic and tool-use workloads, while non-agentic behavior remains largely governed by its inherited base model. Its benefits may be limited for tasks outside the 12 training domains or tool ecosystems unlike those represented in training.
- Palmyra x6 is specialized for agentic/tool-use workloads, with non-agentic behavior largely governed by the inherited base model.
- ∼626 distilled trajectories were used for a single training epoch, bounding data coverage to 12 task domains.
- Tasks far outside these domains or involving substantially different tool ecosystems may not directly benefit.
9 Conclusion
Palmyra x6 is presented as WRITER’s agentic model, post-trained from the GLM 5.2 Mixture-of-Experts base with Anchored Supervised Fine-Tuning on verified synthetic tool-use trajectories. Its deliberately conservative recipe uses 626 trajectories, one epoch, a low learning rate, and a KL anchor to the frozen base, optimized with a Muon + Adam hybrid.
- 9 Conclusion: Palmyra x6 is WRITER’s agentic model, built by post-training the GLM 5.2 Mixture-of-Experts base model.The model uses Anchored Supervised Fine-Tuning on a compact corpus of verified, synthetic tool-use trajectories.
- 9 Conclusion: 626 trajectories, one epoch, a low learning rate, and a KL anchor to the frozen base define the deliberately conservative recipe.The training process is also described as deliberately controlled.
- 9 Conclusion: Muon + Adam hybrid optimization is used for the Anchored Supervised Fine-Tuning procedure.The optimization is applied to the compact corpus of verified, synthetic tool-use trajectories.