|
--- |
|
license: apache-2.0 |
|
task_categories: |
|
- text-generation |
|
--- |
|
Official implementation of the paper ["Order Matters: Investigate the Position Bias in Multi-constraint Instruction Following"](https://huggingface.co/papers/2502.17204). |
|
|
|
Code: https://github.com/meowpass/PBIF |
|
|
|
We systematically study **the position bias problem in multi-constraint |
|
instruction following**. Through our experiments, we have the following findings: |
|
|
|
- **LLMs prefer to "hard-to-easy" constraint order** |
|
- existing LLMs can achieve a better following accuracy in multi-constraint instructions when presented with constraints in “hard-to-easy” orders. |
|
- This finding can be generalized in both single-round and multi-round scenarios, regardless of the architecture of LLM, the size of LLM’s parameters and the number of constraints. |
|
- **Constraints order affect how the LLMs handle a specific constraint** |
|
- The "Hard-to-easy" constraint order induces the LLM to pay more attention to the constraint part in the multi-constraint instructions. |
|
- The LLM’s performance on various constraints is strongly correlated with its attention patterns. |
|
|
|
# PBIF Dataset |
|
The dataset consists of single_round inference data and multi_round inference data. |
|
For each of the data, there are 5 fields: |
|
- `prompt`: Synthesized multi-constraint instructions. |
|
- `constraint`: The constraints contained in the instructions. |
|
- `instruction_id_list`: The id of the constraints in the instructions. |
|
- `kwargs`: Corresponding parameters for the constraints, which are only used for evaluation. |
|
- `ranking`: The constraint order of the instruction. (0 indicates the hardest constraint) |
|
It is worth noting that, in multi_round inference data, the `prompt` is the initial instruction, which is more convenient for the user to construct the multi-round dialog data for themselves. |
