Skip to content

Latest commit

 

History

History
313 lines (253 loc) · 20.3 KB

README.md

File metadata and controls

313 lines (253 loc) · 20.3 KB
Raw

logo

Home | Paper (COLM2024) | Blog | @Maitrix.org

LLM Reasoners is a library to enable LLMs to conduct complex reasoning, with advanced reasoning algorithms. It approaches multi-step reasoning as planning and searches for the optimal reasoning chain, which achieves the best balance of exploration vs exploitation with the idea of "World Model" and "Reward".

Given any reasoning problem, simply define the reward function and an optional world model (explained below), and let LLM reasoners take care of the rest, including Reasoning Algorithms, Visualization, LLM calling, and more!

News

  • Dec. 20, 2024: We now supported planning algorithms (MCTS, DFS/BFS, Beam Search) in web environments with BrowserGym, check the README to try out!

  • Nov. 13, 2024: We integrated DRPO, a tuning-free alignment method published at EMNLP 2024 (link).

  • Jul. 10, 2024: Our paper on LLM Reasoners is accepted to COLM 2024!

  • Jun. 24, 2024: PromptAgent is in LLM Reasoners! Let it help you write down a super detailed prompt for your task (here).

  • May. 14, 2024: Check out Eurus, a suit of LLMs optimized for reasoning. With LLM Reasoners, Eurus-RM can easily boost Llama-8B from 0.49 to 0.73 📈 on GSM8k (code).

  • May. 2, 2024: We have integrated our first reasoning method for scientific reasoning, StructChem! Check it out here.

  • Apr. 22, 2024: We integrated Llama-3, with additional useful APIs (e.g., customizing EOS tokens, calculating likelihood)

  • Apr. 8, 2024: Our new paper introducing LLM Reasoners is available!

  • Mar. 29, 2024: Grace Decoding has been incoporated!

  • Oct. 25, 2023: A video tutorial on the visualizer of LLM Reasoners are available.

  • Oct. 23, 2023: Reasoning-via-Planning is accepted to EMNLP 2023! Check our paper with updated results and discussion!

  • Aug. 21, 2023: A batch of quantized Llama-2 models has arrived! BitsandBytes with huggingface API, GPT-Q with exllama are available. Now you can try llama-2-70B with 2 x 24G GPUs.

  • Aug. 10, 2023: Llama-2 is supported! You can run examples with Llama-2 now.

Why Choose LLM Reasoners?

Experiment Results

  • LLM Reasoners is applied to analyze the reasoning abilities of LLMs and the performance of multiple reasoning algorithms. See the comprehensive experiment results in the AutoRace Leaderboard, and more analysis in the blog and paper.

  • It has been tested to successfully reproduce the performance of Tree-of-Thoughts, Guided Decoding and GRACE Decoding with their official implementation. We list the results reported in their paper / reproduced from their official repositories for reference (†). Some results are on the subsets of the first 100 examples (*).

Method Base LLM GSM8k
Guided Decoding† CodeX (PAL) 0.80
Guided Decoding CodeX (PAL) 0.83*
Method Base LLM Game of 24
Tree-of-Thoughts† GPT-3.5-turbo 0.22
Tree-of-Thoughts GPT-3.5-turbo 0.22
Method Base LLM GSM8k
GRACE Decoding† Flan-T5-Large (Fine-tuned) 0.34
GRACE Decoding Flan-T5-Large (Fine-tuned) 0.33*

Background of LLM Reasoning

Consider the following problem:

Alt text

Let's start with a naive method for LLM reasoning: Prompted with a few examples of problem-solving step by step, an LLM can generate a chain of thoughts (or a sequence of actions) to solve a new problem. For the problem above, the prompt inputted to the LLM and the expected output (in bold) is shown below:

I am playing with a set of blocks where I need to arrange the blocks into stacks.

(Example problems and solutions * 4)

[STATEMENT] 
As initial conditions I have that, the red block is clear, the blue block is clear, the orange block is clear, the hand is empty, the red block is on the yellow block, the yellow block is on the table, the blue block is on the table and the orange block is on the table. My goal is to have that the orange block is on top of the blue block and the yellow block on top of the orange block.

[PLAN]
pick up the orange block
stack the orange block on top of the blue block
unstack the red block from on top of the yellow block
put the red block on the table
pick up the yellow block
stack the yellow block on top of the orange block

Regarding each reasoning step as an action, we have $a_1=$"pick up the orange block", $a_2=$"stack the orange block on top of the blue block", and so on. At each time step, the next action is sampled from the LLM conditioned on the previous actions. This simple method is often referred to as Chain-of-thoughts reasoning. Unfortunately, it doesn't always work for complex reasoning problems. For Blocksworld dataset where the problem above comes from, even the strongest GPT-4 model can only reach the success rate of ~30%.

LLM Reasoners formulate reasoning as planning (RAP). Different from Chain-of-thoughts reasoning which autoregressively samples the next action, our goal is to efficiently search in the reasoning space for the optimal reasoning chain. To achieve this, two components need to be defined: a world model and a reward function.

  • World model defines the state transition, formally $P(s_{i+1} | s_i, a_i)$. A default world model regards the partial solution as the state and simply appends a new action/thought to the state as the state transition (the same formulation of Tree-of-Thoughts). However, you’ll have the option to design a better world model which predicts and keeps track of a more meaningful state (e.g., environment status, intermediate variable values, etc. Check RAP for more examples), thus enhancing the reasoning. For the example shown above, we can naturally define the state as the condition of blocks (e.g., the red block is on the yellow block...), and a world model is to predict the condition of blocks after every potential action.

  • Reward function provides a criterion to evaluate a reasoning step. Ideally, a reasoning chain with a higher accumulated reward should be more likely to be correct. For the example shown above, we can reward actions based on the increased number of accomplished subgoals they lead to. Besides, the likelihood of LLMs generating the action can also be used as a reward, to give the search a good prior.

After we have the world model and reward function, it's time to apply an algorithm to search for the optimal reasoning trace. Here, we show the process of Monte-Carlo Tree Search with a gif:

MCTS Animation

Introduction of the library

The three key components in a reasoning algorithm, reward function, world model, and search algorithm in the formulation (top), correspond to three classes in the library, SearchConfig, WorldModel and SearchAlgorithm respectively. Besides, there are LLM APIs to power other modules, Benchmark, and Visualization to evaluate or debug the reasoning algorithm (middle). To implement a reasoning algorithm for a certain domain (a Reasoner object), a user may inherit the SearchConfig and WorldModel class, and import a pre-implemented SearchAlgorithm. We also show a concrete example of solving Blocksworld with RAP using LLM Reasoners (bottom).

Library Structure

Quick Tour

Let's go through the code of reasoning over Blocksworld problems. Note that the code is simplified for demonstration (check here for a runnable notebook).

The first step is to define the world model: you will set up an initial state given a question in init_state, judge whether a state is terminal in is_terminal, and most importantly, define the world dynamics with step:

from typing import NamedTuple
import utils
from reasoners import WorldModel, LanguageModel
import copy

BWState = str
BWAction = str

class BlocksWorldModel(WorldModel[BWState, BWAction]):
    def __init__(self,
                 base_model: LanguageModel,
                 prompt: dict) -> None:
        super().__init__()
        self.base_model = base_model
        self.prompt = prompt

    def init_state(self) -> BWState:
        # extract the statement from a given problem
        # e.g., "the red block is clear, the blue block is clear..."
        return BWState(utils.extract_init_state(self.example)) 

    def step(self, state: BWState, action: BWAction) -> tuple[BWState, dict]:
        # call the LLM to predict the state transition
        state = copy.deepcopy(state)
        # load the prompt for the LLM to predict the next state
        # e.g. "... I have that <state>, if I <action>, then ..."
        world_update_prompt = self.prompt["update"].replace("<state>", state).replace("<action>", action)
        world_output = self.base_model.generate([world_update_prompt],
                                    eos_token_id="\n", hide_input=True, temperature=0).text[0].strip()
        new_state = utils.process_new_state(world_output)
        # till now, we have the new state after the action
        # the following part is to speed up the reward calculation

        # we want to check the portion of the satisfied subgoals, and use it as a part of the reward
        # since we have predicted the new state already, we can just check it here at convenience
        goal_reached = utils.goal_check(utils.extract_goals(self.example, new_state))
        # return the new state and the additional dictionary (to be passed to the reward function)
        return new_state, {"goal_reached": goal_reached}

    def is_terminal(self, state: BWState) -> bool:
        # define the condition the terminal state to stop the search
        # e.g., all the subgoals are met
        if utils.goal_check(utils.extract_goals(self.example), state.blocks_state) == 1:
            return True
        return False

Then, it's time to consider how to search for the optimal reasoning chain. It involves get_actions to get the action space given a state, and the most important reward as the guidance for reasoning. For Monte-Carlo Tree Search, we can additionally define a fast_reward to speed up the roll-out stage.

import utils
from world_model import BWState, BWAction
from reasoners import SearchConfig, LanguageModel
class BWConfig(SearchConfig):
    def __init__(self,
                 base_model: LanguageModel,
                 prompt: dict,
                 reward_alpha=0.5,
                 goal_reward_default=0.,
                 goal_reached_reward=100) -> None:
        super().__init__()
        self.base_model = base_model
        self.example = None
        self.prompt = prompt
        # some parameters to calculate the fast reward or reward (explained below)
        self.reward_alpha = reward_alpha
        self.goal_reward_default = goal_reward_default
        self.goal_reached_reward = goal_reached_reward

    def get_actions(self, state: BWState) -> list[BWAction]:
        # use a rule-based function to extract all legal actions
        return utils.generate_all_actions(state)

    def fast_reward(self, state: BWState, action: BWAction) -> tuple[float, dict]:
        # build an in-context learning prompt (similar to the one used in Chain-of-thoughts reasoning)
        inputs = self.prompt["icl"].replace("<init_state>", state)\
            .replace("<goals>", utils.extract_goals(self.example))
        # concatenate a candidate action after the prompt, and test its loglikelihood
        intuition = self.base_model.get_loglikelihood(inputs, [inputs + action])[0]
        # the reward is a combination of intuition and goal satisfaction
        # in fast_reward, we skip the calculation of goal satisfaction and use a default value
        fast_reward = intuition * self.reward_alpha + self.goal_reward_default * (1 - self.reward_alpha)
        # cache some information for the reward calculation later (will be passed to `reward` function)
        details = {'intuition': intuition}
        return fast_reward, details

    def reward(self, state: BWState, action: BWAction,
               intuition: float = None,
               goal_reached: tuple[bool, float] = None) -> float:
        # note that `intuition` (cached in `fast_reward`) and `goal_reached` (cached in `step`) are automatically passed as parameters to this reward function
        if goal_reached == 1:
            # if the goal state is reached, we will assign a large reward
            goal_reward = self.goal_reached_reward
        else:
            # otherwise assign the reward based on the portion of satisfied subgoals
            goal_reward = goal_reached
        # the reward is a combination of intuition and goal satisfaction
        reward = intuition * self.reward_alpha + goal_reward * (1 - self.reward_alpha)
        # return the reward and an additional dictionary (to be saved in the log for visualization later)
        return reward, {'intuition': intuition, 'goal_reached': goal_reached}

Now, we are ready to apply a reasoning algorithm to solve the problem:

from reasoners.algorithm import MCTS
from reasoners.lm import LLaMAModel
from world_model import BlocksWorldModel
from search_config import BWConfig

llama_model = LLaMAModel(llama_ckpts, llama_size, max_batch_size=1)
with open(prompt_path) as f:
    prompt = json.load(f)
world_model = BlocksWorldModel(base_model=base_model, prompt=prompt)
config = BWConfig(base_model=llama_model, prompt=prompt)
# save the history of every iteration for visualization
search_algo = MCTS(output_trace_in_each_iter=True)
reasoner = Reasoner(world_model=world_model, search_config=config, search_algo=search_algo)
for i, example in enumerate(dataset):
    algo_output = reasoner(example)
    # save the MCTS results as pickle files
    with open(os.path.join(log_dir, 'algo_output', f'{resume + i + 1}.pkl'), 'wb') as f:
        pickle.dump(algo_output, f)

Finally, we can easily visualize the reasoning process:

import pickle
from reasoners.visualization import visualize
with open("logs/bw_MCTS/xxx/algo_output/1.pkl", 'rb') as f:
    mcts_result = pickle.load(f)

from reasoners.visualization.tree_snapshot import NodeData
from reasoners.algorithm.mcts import MCTSNode

# by default, a state will be presented along with the node, and the reward with saved dictionary in `SearchConfig.reward` will be presented along with the edge. 
# we can also define a helper function to customize what we want to see in the visualizer.
def blocksworld_node_data_factory(n: MCTSNode) -> NodeData:
    return NodeData({"block state": n.state.blocks_state if n.state else None,
                     "satisfied": n.fast_reward_details if n.fast_reward_details else "Not expanded"})
def blocksworld_edge_data_factory(n: MCTSNode) -> EdgeData:
    return EdgeData({"reward": n.reward, "intuition": n.fast_reward_details["intuition"]})
visualize(mcts_result, node_data_factory=blocksworld_node_data_factory,
                       edge_data_factory=blocksworld_edge_data_factory)

Then a URL of the visualized results will pop up. The figure will be interactive and look like the examples shown on our demo website.

Installation

Make sure to use Python 3.10 or later.

conda create -n reasoners python=3.10
conda activate reasoners

Install from pip

pip install llm-reasoners

Install from github

(Recommended if you want to run the examples in the github repo)

git clone https://github.com/Ber666/llm-reasoners --recursive
cd llm-reasoners
pip install -e .

Adding --recursive will help you clone exllama and LLM-Planning automatically. Note that some other optional modules may require other dependencies. Please refer to the error message for details.

Citation

This project is an extension of the following paper:

@inproceedings{hao2023reasoning,
  title={Reasoning with Language Model is Planning with World Model},
  author={Hao, Shibo and Gu, Yi and Ma, Haodi and Hong, Joshua and Wang, Zhen and Wang, Daisy and Hu, Zhiting},
  booktitle={Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing},
  pages={8154--8173},
  year={2023}
}
@article{hao2024llm,
  title={LLM Reasoners: New Evaluation, Library, and Analysis of Step-by-Step Reasoning with Large Language Models},
  author={Hao, Shibo and Gu, Yi and Luo, Haotian and Liu, Tianyang and Shao, Xiyan and Wang, Xinyuan and Xie, Shuhua and Ma, Haodi and Samavedhi, Adithya and Gao, Qiyue and others},
  journal={arXiv preprint arXiv:2404.05221},
  year={2024}
}