DEVOURED
Mammoth-1 and Elephant-1: frontier decision models
Base Labs released Mammoth-1 and Elephant-1, two specialized decision models that outperform general-purpose models on decision-specific benchmarks by using block-bidirectional attention.
What: Mammoth-1 (27B) and Elephant-1 (9B) are fine-tuned Qwen3.5 models optimized for probability distributions over choices rather than text generation. Mammoth-1 leads the Jev Decision Index at 71.24.
Why it matters: This demonstrates a shift toward specialized model architectures—decoupling decision-making from fluent text generation—to improve certainty calibration and reduce inference costs for enterprise tasks.
Takeaway: If you are currently using LLMs for classification or routing tasks, testing decision-specific models like Mammoth-1 may yield higher accuracy and lower latency.
Deep dive
- Mammoth-1 and Elephant-1 are fine-tuned versions of Qwen3.5-27B and 9B respectively.
- Architecture replaces standard causal attention masks with block-bidirectional masks for full-attention layers.
- This allows questions to attend to their full option set while preventing the system prompt/state from attending to the question itself.
- Models are trained to return probability distributions across options rather than generating text.
- The training pipeline uses synthetic data generation to address model failure modes, growing the mixture to 6.21M examples.
- Incorporates a Brier term in the loss function to penalize overconfidence and improve calibration.
- The model skips decoding entirely, making inference cheaper and faster than standard autoregressive LLMs.
Decoder
- Decision model: A model trained specifically to output probability distributions over a set of provided options, rather than generating natural language text.
- Brier term: A scoring rule used to measure the accuracy of probabilistic predictions, penalizing the mean squared difference between predicted probability and actual outcome.
- Block-bidirectional attention: An attention mechanism variant where tokens within a defined block can attend to all other tokens in that same block, while maintaining causal (left-to-right) constraints between blocks.
Original article
Full article content is not available for inline reading.
Explore all 58 articles in today's digest
Fresh dev news, AI-digested daily. Never miss a trend at devoured.fyi