Clipped Importance Sampling Policy Optimization (CISPO)¶

CISPO (Clipped Importance Sampling Policy Optimization) is a GRPO specialization that clips importance weights directly and uses them to scale a log-prob objective.

CISPO uses the same group-based advantage calculation as GRPO, however, the objective function is closer to that of REINFORCE, multiplying the log-probability term of the function by a scaled importance ratio. A stop gradient is applied to the importance ratio, meaning the ratio is treated as a constant that scales each token’s contribution to the overall policy gradient.

In AgileRL, CISPO can be used for single-turn reasoning tasks or multi-turn agentic finetuning. In the multi-turn case, rollouts are still treated as a bandit problem, with environment generated tokens masked and reward signal calculated from cumulative episode reward.

Example¶

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
from agilerl.algorithms import CISPO

model = AutoModelForCausalLM.from_pretrained(
    "Qwen/Qwen2.5-3B",
    torch_dtype=torch.bfloat16,
    device_map="auto",
)
tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen2.5-3B")

agent = CISPO(
    actor_network=model,
    pad_token_id=tokenizer.eos_token_id,
    pad_token=tokenizer.eos_token,
    device="cuda" if torch.cuda.is_available() else "cpu",
    batch_size=8,
    group_size=8,
)

Training and Usage¶

Use CISPO anywhere you would use GRPO in AgileRL training loops, such as finetune_llm_reasoning or finetune_llm_multiturn.

from datasets import Dataset
from agilerl.llm_envs import ReasoningGym, TokenObservationWrapper
from agilerl.training.train_llm import (
    finetune_llm_multiturn,
    finetune_llm_reasoning,
)

# 1) Single-turn / reasoning datasets (ReasoningGym)
train_ds = Dataset.from_dict(
    {
        "question": ["2+2?", "Capital of France?"],
        "answer": ["4", "Paris"],
    }
)
test_ds = Dataset.from_dict(
    {
        "question": ["3+3?"],
        "answer": ["6"],
    }
)

def reward_fn(completion: str, answer: str, question: str) -> float:
    del question
    return float(answer.lower() in completion.lower())

reasoning_env = ReasoningGym(
    train_dataset=train_ds,
    test_dataset=test_ds,
    tokenizer=tokenizer,
    reward_fn=reward_fn,
    conversation_template=[{"role": "user", "content": "Q: {question}\nA:"}],
    data_batch_size_per_gpu=2,
)

trained_pop = finetune_llm_reasoning(
    pop=[agent],
    env=reasoning_env,
    max_steps=2000,
    evaluation_interval=50,
)

# 2) Multi-turn text environments (factory + wrapper)
class ToyMultiTurnEnv:
    def reset(self, seed=None):
        del seed
        return "Start: What is 2+2?", {}

    def step(self, action: str):
        reward = 1.0 if "4" in action else 0.0
        return "Done.", reward, True, False, {"correct": bool(reward)}

def env_factory():
    return TokenObservationWrapper(
        env=ToyMultiTurnEnv(),
        tokenizer=tokenizer,
        max_turns=4,
        pad_id=tokenizer.eos_token_id,
        max_model_len=1024,
        max_output_tokens=128,
    )

trained_pop = finetune_llm_multiturn(
    pop=[agent],
    max_turns=4,
    env_factory=env_factory,
    max_steps=2000,
    evaluation_interval=50,
)

Saving and Loading Agents¶

To save an agent, use the save_llm_checkpoint function:

from agilerl.utils.utils import save_llm_checkpoint

save_llm_checkpoint(agent, "path/to/checkpoint")

Parameters¶

class agilerl.algorithms.cispo.CISPO(pad_token_id: int, pad_token: str, model_name: str | None = None, actor_network: PreTrainedModelProtocol | None = None, model_config: dict[str, Any] | None = None, hp_config: HyperparameterConfig | None = None, index: int = 0, batch_size: int = 16, beta: float = 0.001, lr: float = 5e-07, clip_coef: float | tuple[float, float] = 0.2, max_grad_norm: float = 0.1, update_epochs: int = 1, group_size: int = 8, temperature: float = 0.9, repetition_penalty: float = 1.0, top_p: float = 0.95, top_k: int = 50, min_p: float = 0.0, use_memory_efficient_params: bool = True, calc_position_embeddings: bool = True, micro_batch_size_per_gpu: int | None = None, max_output_tokens: int | None = None, min_output_tokens: int | None = None, max_model_len: int | None = 1024, hf_generate_chunk_size: int | None = None, lora_config: LoraConfig | None = None, cosine_lr_schedule_config: CosineLRScheduleConfig | None = None, accelerator: Accelerator | None = None, device: str = 'cpu', wrap: bool = True, clone: bool = False, use_vllm: bool = False, vllm_config: VLLMConfig | None = None, seed: int = 42, gradient_checkpointing: bool = True, torch_compiler: str | None = None, use_liger_loss: bool = False, use_kl_advantage_shaping: bool = False, adv_norm: str = 'mean_std', use_separate_reference_adapter: bool = True, whiten_advantages: bool = False, adv_clip_range: float | None = None, filter_zero_adv: bool = False, adv_filter_eps: float = 0.0, reduce_memory_peak: bool = False, use_fused_linear_logprobs: bool = False, cast_logprobs_to_fp32: bool = True)¶

CISPO loss variant of agilerl.algorithms.grpo.GRPO

Paper: https://arxiv.org/abs/2506.13585

clean_up() → None¶: Clean up the algorithm.

clone(index: int | None = None, wrap: bool = True) → Self¶

Create a clone of the algorithm.

Parameters:

index (int | None, optional) – The index of the clone, defaults to None
wrap (bool, optional) – If True, wrap the models in the clone with the accelerator, defaults to False

Returns:

A clone of the algorithm

Return type:

EvolvableAlgorithm

static copy_attributes(agent: EvolvableAlgorithm, clone: EvolvableAlgorithm) → EvolvableAlgorithm¶

Copy the non-evolvable attributes of the algorithm to a clone.

Parameters:: clone (EvolvableAlgorithm) – The clone of the algorithm.
Returns:: The clone of the algorithm.
Return type:: EvolvableAlgorithm

evolvable_attributes(networks_only: bool = False) → dict[str, EvolvableModuleProtocol | ModuleDictProtocol | Optimizer | dict[str, Optimizer] | OptimizerWrapperProtocol]¶

Return the attributes related to the evolvable networks in the algorithm. Includes attributes that are either EvolvableModule or ModuleDict objects, as well as the optimizers associated with the networks.

Parameters:: networks_only (bool, optional) – If True, only include evolvable networks, defaults to False
Returns:: A dictionary of network attributes.
Return type:: dict[str, Any]

get_action(obs: list[ReasoningPrompts] | ReasoningPrompts, training: bool = True, repeat_prompts: bool = True, *args: Any, **kwargs: Any) → tuple[list[Tensor], list[Tensor]]¶

Return generated completions for each prompt (GRPO groups when training).

Parameters:

obs (LLMObsType) – List of HF-style prompt dicts (this implementation mutates them).
training (bool) – If True, generate with training sampling settings.
repeat_prompts (bool) – If True and training=True, duplicate each prompt self.group_size times (legacy GRPO grouped mode). If False, treat the batch as already expanded trajectories.

Returns:

Completion token IDs and per-sequence action masks.

Return type:

tuple[list[torch.Tensor], list[torch.Tensor]]

Return the dimension of the action space as it pertains to the underlying networks (i.e. the output size of the networks).

Parameters:: action_space (spaces.Space or list[spaces.Space].) – The action space of the environment.
Returns:: The dimension of the action space.
Return type:: int.

get_lr_names() → list[str]¶: Return the learning rates of the algorithm.

get_policy() → EvolvableModuleProtocol¶: Return the policy network of the algorithm.

Return the dimension of the state space as it pertains to the underlying networks (i.e. the input size of the networks).

Parameters:: observation_space (spaces.Space or list[spaces.Space].) – The observation space of the environment.
Returns:: The dimension of the state space.
Return type:: tuple[int, …].

property index: int¶: Return the index of the algorithm.

static inspect_attributes(agent: EvolvableAlgorithm, input_args_only: bool = False) → dict[str, Any]¶

Inspect and retrieve the attributes of the current object, excluding attributes related to the underlying evolvable networks (i.e. EvolvableModule, torch.optim.Optimizer) and with an option to include only the attributes that are input arguments to the constructor.

Parameters:: input_args_only (bool) – If True, only include attributes that are input arguments to the constructor. Defaults to False.
Returns:: A dictionary of attribute names and their values.
Return type:: dict[str, Any]

Update agent network parameters to learn from experiences.

Parameters:: experiences (ExperiencesType) – (completion_ids, action_masks, rewards) stacked batch.
Returns:: Dict with keys mean_loss and mean_kl, averaged over the update.
Return type:: dict[str, float]

classmethod load(path: str, device: str | device = 'cpu', accelerator: Accelerator | None = None) → None¶

Load an algorithm from a checkpoint.

Parameters:

path (string) – Location to load checkpoint from.
device (str, optional) – Device to load the algorithm on, defaults to ‘cpu’
accelerator (Accelerator | None, optional) – Accelerator object for distributed computing, defaults to None

Returns:

An instance of the algorithm

Return type:

RLAlgorithm

load_checkpoint(path: str, load_optimizer: bool = False, overwrite_reference_adapter: bool = False, overwrite_critic_adapter: bool = True, merge_lora_configs: bool = False) → None¶

Load adapter weights and algorithm state from a checkpoint directory.

Adapter roles restored on load:

actor — the trained policy. Always loaded.

reference — the fixed policy used for KL / comparison. The checkpoint’s actor adapter is copied onto reference so that SFT -> DPO -> GRPO chains work out of the box: the stage-N actor becomes the stage-N+1 reference.

critic — optional value head. Loaded from disk if a critic/ adapter is present, else copied from actor, else left as the live fresh LoRA init.

LoRA config reconciliation: when the checkpoint’s config and the live algorithm’s config disagree, loading fails fast by default. Pass merge_lora_configs=True to merge them for compatibility:

r (rank) -> max(current, checkpoint); the smaller side’s weights are padded into the top-left rank slice of the larger adapter (see _pad_adapter_state_to_live_shape()).

target_modules / modules_to_save -> union.

Any other mismatched field -> current value wins, with a warning.

Any adapter whose live config ends up differing from the selected target config is rebuilt via _reconfigure_adapters_to_match() before weights are loaded, so tensors always land in the correct shape.

No DeepSpeed:

lora_only=T, load_optimizer=T -> PEFT adapter load + optimizer
state from attributes.pt

lora_only=T, load_optimizer=F -> PEFT adapter load only lora_only=F, load_optimizer=T -> torch load of actor +

optimizer from attributes.pt

lora_only=F, load_optimizer=F -> torch load of actor only

DeepSpeed:

lora_only=T, load_optimizer=T -> DeepSpeed engine load from
<path>/save_checkpoint

lora_only=T, load_optimizer=F -> PEFT adapter load lora_only=F, load_optimizer=T -> DeepSpeed engine load from

<path>/save_checkpoint

lora_only=F, load_optimizer=F -> actor.load_state_dict(...)
from attributes.pt

When load_optimizer=True but the checkpoint contains no optimizer state (e.g. it was saved with save_optimizer=False), a UserWarning is emitted and a freshly-initialised optimizer is used.

Parameters:

path (str) – Directory containing a checkpoint written by save_checkpoint().
load_optimizer (bool) – If True (default) also load the optimizer and LR scheduler state so training can resume. On DeepSpeed ZeRO ≥ 2 this reads a sharded checkpoint from <path>/save_checkpoint; otherwise optimizer state is read from attributes.pt.
merge_lora_configs (bool) – If True, allow loading checkpoints whose LoRA config differs from the live agent by reconciling them. If False (default), mismatched LoRA configs raise ValueError.

property mut: Any¶: Return the mutation object of the algorithm.

mutation_hook() → None¶: Execute the hooks registered with the algorithm.

Create a population of algorithms.

Parameters:: size (int.) – The size of the population.
Returns:: A list of algorithms.
Return type:: list[EvolvableAlgorithm].

Preprocess observations (dummy) for forward pass through neural network.

Parameters:: observations (numpy.ndarray[float] or dict[str, numpy.ndarray[float]]) – Observations of environment
Returns:: Preprocessed observations
Return type:: torch.Tensor[float] or dict[str, torch.Tensor[float]] or tuple[torch.Tensor[float], …]

recompile() → None¶

Recompile evolvable modules with torch.compile.

Iterates over evolvable_attributes and compiles each one. Skipped when DeepSpeed is active because DeepSpeedEngine is not compatible with OptimizedModule wrapping.

register_mutation_hook(hook: Callable) → None¶

Parameters:: hook (Callable) – The hook to be executed after mutation.

register_network_group(group: NetworkGroup) → None¶

Set the evaluation network for the algorithm.

Parameters:: name (str) – The name of the evaluation network.

reinit_optimizers(optimizer: OptimizerConfig | None = None) → None¶

Reinitialize the optimizers of an algorithm. If no optimizer is passed, all optimizers are reinitialized.

Parameters:: optimizer (OptimizerConfig | None, optional) – The optimizer to reinitialize, defaults to None, in which case all optimizers are reinitialized.

save_checkpoint(path: str, lora_only: bool = True, save_optimizer: bool = True, **kwargs: Any) → None¶

Save adapter weights and algorithm state to a directory.

AgileRL never persists base-model weights when lora_only=True for LLM algorithms: a checkpoint is a directory containing

<adapter>/adapter_model.safetensors + adapter_config.json — one subdirectory per adapter in selected_adapters (always actor, plus reference / critic when those adapters are configured). Written only when lora_only=True.

attributes.pt — algorithm hyperparameters, plus (optionally) the actor state dict and/or optimizer state dict depending on the cell below. Always present.

save_checkpoint/ — DeepSpeed ZeRO ≥ 2 sharded-checkpoint output. Present only when an Accelerator is attached and save_optimizer=True.

Behaviour per cell of the (lora_only, save_optimizer, deepspeed) grid:

Plain (no accelerator):

lora_only=T, save_optimizer=T -> PEFT adapter dirs on disk +
optimizer state in attributes.pt

lora_only=T, save_optimizer=F -> PEFT adapter dirs only lora_only=F, save_optimizer=T -> full actor state_dict +

optimizer state in attributes.pt

lora_only=F, save_optimizer=F -> full actor state_dict in attributes.pt

DeepSpeed:

lora_only=T, save_optimizer=T -> engine tag dir (frozen params
excluded) + PEFT adapter dirs

lora_only=T, save_optimizer=F -> PEFT adapter dirs only lora_only=F, save_optimizer=T -> engine tag dir (frozen params

included)

lora_only=F, save_optimizer=F -> gathered (ZeRO-3 aware) actor
state_dict injected into attributes.pt

Parameters:

path (str) – Directory to write the checkpoint into.
lora_only (bool) – If True (default) only adapter weights are written to disk via save_pretrained; the base model is shared across checkpoints and not serialised. If False, the full actor state dict is persisted (into attributes.pt on the plain path, or into the DeepSpeed engine’s tag dir / gathered dict on the distributed path).
save_optimizer (bool) – If True (default) also persist the optimizer and LR scheduler state so training can resume. On DeepSpeed ZeRO ≥ 2 this writes a sharded checkpoint into <path>/save_checkpoint; otherwise optimizer state is included in attributes.pt.

select_adapter(adapter_name: str) → None¶

Temporarily switch adapter; restores the actor adapter on exit.

Parameters:: adapter_name (str) – Name of the adapter to activate (“actor”, “critic”, “reference”).

set_reference_policy(reference_update_tracker: int) → None¶

Update the reference policy when the reference policy update tracker is greater than the current reference policy update tracker.

Parameters:: reference_update_tracker (int) – The reference policy update tracker

set_training_mode(training: bool) → None¶

Set the training mode of the algorithm.

Parameters:: training (bool) – If True, set the algorithm to training mode.

test(env: ReasoningGym | MultiTurnEnv, loop: int = 1, *args: Any, **kwargs: Any) → ndarray¶

Return fitness (test) score of llm on test sub-set.

Parameters:

env (ReasoningGym | MultiTurnEnv) – Dataset-style ReasoningGym environment or tokenized multi-turn episode environment.
loop (int) – Number of outer test iterations over reset / step.

Returns:

Concatenated reward tensor from the test loop.

Return type:

torch.Tensor

Move experiences to the device.

Parameters:: experiences (tuple[torch.Tensor[float], ...]) – Experiences to move to device
Returns:: Experiences on the device
Return type:: tuple[torch.Tensor[float], …]

unwrap_models() → None¶: Unwraps the models in the algorithm from the accelerator.

Update the learning rate of the optimizer.

Parameters:

optimizer (Optimizer) – Optimizer
lr (float | tuple[float, float]) – Learning rate value, or actor/critic pair.
accelerator (Accelerator | None) – Accelerator
scheduler_config (CosineLRScheduleConfig | None) – Scheduler configuration

Returns:

Tuple of accelerator and scheduler

Returns:

Accelerator

use_adapter(adapter_name: str) → None¶

Switch the active PEFT adapter, handling all side-effects.

For “reference”: switches adapter and freezes reference params (never trained). For all others: switches adapter and restores requires_grad=True on all training adapter LoRA params so that DeepSpeed ZeRO-2 gradient bucket hooks keep firing correctly.

Parameters:: adapter_name (str) – Name of the adapter to activate (“actor”, “critic”, “reference”).

wrap_models() → None¶: Wrap the models in the accelerator, DeepSpeed objects must be wrapped at the same time, not individually.