TL;DR
JT-Safe-V2 introduces a "safety-by-design" paradigm for large language models by embedding a three-layer "World Context" directly into pre-training data and utilizing a deterministic, high-certainty training procedure. Alongside this safety-enhanced 35B foundation model, the authors present Safe-MoMA, an intelligent multi-agent orchestration framework that slashes enterprise inference costs by over 30% while maintaining top-tier performance.
What Problem Does This Paper Address?
Traditional LLM safety strategies operate like an external guardrail system applied after a model is fully trained (e.g., via RLHF or safety classifiers). However, many reliability issues—such as hallucinations, factual contradictions, and vulnerability to adversarial prompts—originate much earlier in the pipeline.
Standard web-scale pre-training corpora consist of "flattened" text that lacks temporal context, author intent, or causal structures. When models train purely on next-token prediction over this ambiguous data, they internalize uncertainty. JT-Safe-V2 addresses this by shifting safety from a reactive post-training patch to an intrinsic design objective embedded across the entire model lifecycle: data construction, pre-training, and alignment.
Methods
The authors introduce several core innovations to realize their safety-by-design paradigm:
- Data with World Context (DWC): Rather than treating text as isolated token sequences, the DWC framework enriches data with a structured, three-layer annotation architecture:
- Factual Layer: Anchors text with explicit spatiotemporal metadata, domain classifications, and source credibility indicators.
- Logical Layer: Exposes underlying reasoning structures, including causal relations, conditional dependencies, and chain-of-thought traces.
-
Cognitive Layer: Captures the informational intent, reading purpose, and value perspectives.
Figure 1: The three-layer annotation architecture of DWC. -
High-Certainty Pre-training (WSM Paradigm): Traditional cosine learning rate decay schedules can cause optimization instability. The Warmup-Stable-Merge (WSM) paradigm decouples parameter optimization from learning rate annealing. The learning rate is held at a high constant value for broad parameter space exploration, and checkpoints are merged offline using diversified weight distribution strategies. This prevents catastrophic forgetting during continual pre-training.
Figure 1: The three-layer annotation architecture of DWC. -
Prefix-Guided Meta-Information Activation: To prevent the model from failing to elicit its internal safety knowledge when given concise user queries, the authors use post-training fine-tuning to map quality and safety constraints to specific system prompt prefixes. These prefixes act as triggers to awaken latent DWC knowledge during inference.
- Safe-MoMA (Safe Mixture of Models and Agents): A task-driven orchestration framework formulated as a Markov Decision Process (MDP) and optimized via reinforcement learning. It features a high-level routing policy that selects the best solving paradigm (single model, single agent, or multi-agent) and a low-level policy that dynamically allocates specific models and tools based on task complexity and resource constraints.
Datasets
- Pre-training & Alignment: The model utilizes the newly developed DWC corpus and a seed cold-start dataset of approximately $10^5$ high-quality instruction samples for model-driven self-distillation.
- Safety Benchmarks: Evaluated across 20 benchmarks spanning six categories, including CNsafe, CValues, Flames, Sweeval, Airbench, AgentHarm, SafetyPrompts, JADE, SimpleSafetyTest, Forbidden, DoNotAnswer, TechHazardQA, SaladBench, BBQ, JailbreakBench, StrongReject, Jbdistill, CSSbench, ChineseSafetyQA, TruthfulQA, and RagTruth.
- General Capabilities Benchmarks: Evaluated using standard benchmarks including HumanEval(+), MBPP+, CruxEval, MultiPL-E (coding); GSM8K, MATH500, MGSM-zh, MathBench (math); MRCR, RULER, Leval (long context); PIQA, HellaSwag, StrategyQA, AutoLogi (reasoning); CryptoBench, FinEval, Xiezhi, GaokaoBench (knowledge); and BFCL, $\tau^2$-Bench, SEAL, GAIA-text (agentic capabilities).
Key Insights and Ideas
- Safety Enhances Capability: A common assumption is that strict safety alignment degrades an LLM's general capabilities. JT-Safe-V2 proves the opposite: structuring pre-training data with factual, logical, and cognitive context layers actually boosts performance in downstream reasoning, knowledge retrieval, and long-context tasks.
- Plug-and-Play Meta-Information: The DWC meta-information exhibits strong plug-and-play compatibility. It can be injected into an already well-trained checkpoint during continued pre-training without requiring architectural changes or re-initialization, significantly accelerating data efficiency.
- Dynamic Orchestration Over Monolithic Scaling: Deploying a single massive model for every enterprise task is highly inefficient. By framing multi-model routing as a hierarchical reinforcement learning problem, systems can achieve state-of-the-art performance by dynamically combining lightweight models, specialized agents, and heavy reasoning models only when necessary.
Key Results
- Safety Dominance: JT-Safe-V2-35B achieved a near-perfect score of 100.00 on CValues, 99.66 on CNsafe, and 99.79 on StrongReject, consistently outperforming or matching much larger models like Qwen3-235B and DeepSeek-V3.2.
- General Capabilities: The safety-first training did not degrade performance; the model achieved 96.30 on HumanEval (compared to the 94.50 baseline SOTA) and dramatically improved on GaokaoBench (92.28 vs. 27.76) and MathBench (94.87 vs. 74.26).
- Safe-MoMA Efficiency: Compared to using the optimal standalone large model (Qwen3-235B-A22B), the Safe-MoMA routing framework achieved a 2.9% score improvement while reducing inference costs by 31.46% under performance-priority settings, and up to 42% cost reduction overall.
Limitations and Future Work
- Agentic Capabilities: While JT-Safe-V2 showed strong planning improvements on $\tau^2$-Bench, it lagged behind on benchmarks like SEAL and GAIA-text. The authors explicitly call out agentic capability as a key area for future improvement.
- Routing in Unconstrained Categories: The SFT-based approach for task routing performs exceptionally well under limited, well-defined data categories, but its performance tends to degrade when applied to highly diverse, unconstrained real-world enterprise scenarios.
Summary updated