SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2403.09472 · NeurIPS · 2024

Easy-to-Hard Generalization: Scalable Alignment Beyond Human Supervision

Chuang Gan, Yiming Yang, Zhiqing Sun, Sean Welleck, Weiyang Liu, Yikang Shen, Longhui Yu

arXiv · PDF · Open in the Atlas

Code that ran

5 of 5 samples (functions and classes) lifted from this paper's linked repositories executed on a synthesized input in Syntology's jail (0 of them constructed an object rather than computing a result; 1 honoured the contract we drafted, 0 violated it and 0 have no contract verdict; 4 failures of our instrument; 0 no recorded run); 2 of them were also run on the host, not in the jail, by the peer-agreement lane that set their level, measured 2026-11-22; repository links from the Papers with Code archive attributed under CC BY-SA, no affiliation; bodies inline only under a permissive licence as GitHub detects it. Not a reproduction of the paper's results.

"Ran" means the function executed on a synthesized input and returned a value. It is not a reproduction of the paper's results. This record, dated, as JSON: /api/ran/2403.09472.json · badge: /api/ran/2403.09472.svg

Each repository's role is a label: a classifier's output recorded on the link, or, where no role is recorded, what the Papers with Code archive marked. The line under it names its basis. Neither says who wrote the code.

RepositoryRoleRan
edward-sun/easy-to-hard canonical (classifier label)
classifier basis: classifier_2025
5 ran · 0 no recorded run
FunctionStatusWhere it lives
batch_select Ran
ran; our contract draft was wrong, not the code
edward-sun/easy-to-hard/models/reward_model.py
code served (permissive licence) · get_code("2154e44a6b59e432")
How it ran

Ran in Syntology's jail on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1) and returned Tensor, ndim 2, dtype family int, shape (2, 3) as a later run recorded it. The image it ran in was not recorded. The comparison behind this sample's level is a different run: the peer-agreement lane ran it again on the host, not in the jail (report dated 2026-09-04T22:51:12Z). If your copy returns the same type, ndim and shape on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch

torch.manual_seed(0)

# Create input tensor: batch_size=2, num_features=4
input = torch.tensor([
    [0, 1, 2, 5],
    [3, 0, 9, 4],
])

# Create index tensor: batch_size=2, num_selections=3
# Each row contains indices to select from the corresponding input row
index = torch.tensor([
    [0, 1, 3],
    [1, 0, 2],
])

ARGS = {
    "input": input,
    "index": index,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type Tensor, ndim 2, dtype family int, shape (2, 3) — shape (2, 3) as a later run recorded it
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
unpack_dict Ran
ran; our contract draft was wrong, not the code
edward-sun/easy-to-hard/models/reward_model.py
code served (permissive licence) · get_code("a6989412aca721c1")
How it ran

Ran in Syntology's jail on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1) and returned tuple. The image it ran in was not recorded. The comparison behind this sample's level is a different run: the peer-agreement lane ran it again on the host, not in the jail (report dated 2026-09-05T01:20:33Z). If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch

torch.manual_seed(0)

# Create a sample dictionary with various keys and values
d = {
    'key1': torch.randn(2, 4),
    'key2': torch.randn(3, 5),
    'key3': 'value3',
    'key4': 42,
    'key5': [1, 2, 3],
}

# Define keys to extract
keys = ['key1', 'key3', 'key4']

# Define return type
return_type = tuple

ARGS = {
    'd': d,
    'keys': keys,
    'return_type': return_type,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type tuple — a tuple output has no array shape of its own; none is stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
compute_pairwise_reward_modeling_loss Ran
ran, honoured the contract we drafted
edward-sun/easy-to-hard/models/reward_model.py
code served (permissive licence) · get_code("36772c05893440aa")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1) and returned Tensor, ndim 0, dtype family float, shape () as a later run recorded it. The image it ran in was not recorded. If your copy returns the same type, ndim and shape on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import torch.nn as nn

torch.manual_seed(0)

# Build a minimal model that satisfies the function's usage
class MinimalRewardModel(nn.Module):
    def __init__(self):
        super().__init__()
        self.linear = nn.Linear(4, 1)
    
    def forward(self, input_ids, input_pos, eos_pos):
        # input_ids: (batch*candidates, seq_len)
        # input_pos: (seq_len,)
        # eos_pos: (batch*candidates,)
        batch_size = input_ids.size(0)
        # Return an object with a rewards attribute
        rewards = self.linear(input_ids.float()).squeeze(-1)  # (batch*candidates,)
        
        class Output:
            pass
        
        output = Output()
        output.rewards = rewards
        return output

model = MinimalRewardModel()

# Construct inputs dict
batch_size = 2
num_candidates = 3
seq_len = 4
num_pairs = 2

input_ids = torch.randint(0, 10, (batch_size, num_candidates, seq_len))
eos_pos = torch.randint(0, seq_len, (batch_size, num_candidates))
index_0 = torch.randint(0, num_candidates, (batch_size, num_pairs))
index_1 = torch.randint(0, num_candidates, (batch_size, num_pairs))
choice = torch.randint(0, 2, (batch_size, num_pairs)).float()

inputs = {
    'input_ids': input_ids,
    'eos_pos': eos_pos,
    'index_0': index_0,
    'index_1': index_1,
    'choice': choice,
}

ARGS = {
    'model': model,
    'inputs': inputs,
    'return_outputs': False,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1
Imports
einops, torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type Tensor, ndim 0, dtype family float, shape () — shape () as a later run recorded it
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
compute_reward_modeling_metrics Ran
ran; our contract draft was wrong, not the code
edward-sun/easy-to-hard/train_rl_dpo.py
code served (permissive licence) · get_code("d4969006e20577f6")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1) and returned dict. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch

torch.manual_seed(0)

# Create small tensors for logits and labels
# logits: continuous values that will be thresholded at 0.0
# labels: binary labels (0 or 1)
logits = torch.randn(2, 4)  # batch_size=2, some features
labels = torch.randint(0, 2, (2, 4)).long()  # binary labels

ARGS = {
    "logits": logits,
    "labels": labels,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type dict — a dict output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a dict output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
model_forward Ran
ran; our contract draft was wrong, not the code
edward-sun/easy-to-hard/train_rl_dpo.py
code served (permissive licence) · get_code("63f447f7130d312a")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1) and returned Tensor, ndim 3, dtype family float, shape (2, 3, 8) as a later run recorded it. The image it ran in was not recorded. If your copy returns the same type, ndim and shape on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import torch.nn as nn

torch.manual_seed(0)

class SimpleModel(nn.Module):
    def __init__(self):
        super().__init__()
        self.linear = nn.Linear(4, 8)
    
    def forward(self, x, input_pos, fully_causal=False):
        # x shape: (batch, seq_len, dim)
        # input_pos: position indices
        # fully_causal: boolean flag
        return self.linear(x)

model = SimpleModel()
x = torch.randn(2, 3, 4)  # batch=2, seq_len=3, dim=4
input_pos = torch.tensor([0, 1, 2])  # position indices

ARGS = {
    "model": model,
    "x": x,
    "input_pos": input_pos,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type Tensor, ndim 3, dtype family float, shape (2, 3, 8) — shape (2, 3, 8) as a later run recorded it
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block

Also present in other papers' code

Of 5 samples lifted from this paper's linked repositories, 5 carry a function identity and 0 do not. Of the 5 distinct functions with an identity, 3 are also present in other papers whose code we hold (2 with at least one paper dated earlier, 0 the same month, 3 later) and 2 are only here among the 52,977 papers whose code Syntology holds (as of 2026-09-26).

Only here means: among the 52,977 papers whose code Syntology holds (51,887 with lifted samples, 1,090 through their assemblies only), as of 2026-09-26, no other paper's lifted samples or current assembly bricks contain an identical function body after normalisation.

FunctionAlso present inEarliest-datedOthers named
batch_select 3 other papers
1 dated earlier, 2 later
2310.05910 (2023-10)
SALMON: Self-Alignment with Instructable Reward Models
2406.11280 (2024-06)
ISR-DPO: Aligning Large Multimodal Models for Videos by Iterative Self-Retrospective DPO
2406.12030 (2024-06)
SPA-VL: A Comprehensive Safety Preference Alignment Dataset for Vision Language Model
unpack_dict 3 other papers
1 dated earlier, 2 later
2310.05910 (2023-10)
SALMON: Self-Alignment with Instructable Reward Models
2406.11280 (2024-06)
ISR-DPO: Aligning Large Multimodal Models for Videos by Iterative Self-Retrospective DPO
2406.12030 (2024-06)
SPA-VL: A Comprehensive Safety Preference Alignment Dataset for Vision Language Model
model_forward 1 other paper
1 later
2408.00724 (2024-08)
Inference Scaling Laws: An Empirical Analysis of Compute-Optimal Inference for Problem-Solving with Language Models

Presence, not direction: an identical function body in two papers' repositories does not say who wrote it, which came from which, or that either paper's method rests on it. Matching is exact after normalisation, so a renamed or edited function is not matched. Dates are when papers appeared, not when code was written. Only .py files the harvester lifted are covered; notebooks are not read. A new-style arXiv id YYMM.NNNN dates a paper to the month 20YY-MM; otherwise Paper.year dates it to the year; otherwise it is undated. Dates are when the paper appeared, not when the code was written. Read 2026-11-22.

Repositories linked to this paper

Some links come from the archived Papers with Code dataset (CC BY-SA 4.0): attribution and licence.

Abstract

Current AI alignment methodologies rely on human-provided demonstrations or judgments, and the learned capabilities of AI systems would be upper-bounded by human capabilities as a result. This raises a challenging research question: How can we keep improving the systems when their capabilities have surpassed the levels of humans? This paper answers this question in the context of tackling hard reasoning tasks (e.g., level 4-5 MATH problems) via learning from human annotations on easier tasks (e.g., level 1-3 MATH problems), which we term as easy-to-hard generalization. Our key insight is that an evaluator (reward model) trained on supervisions for easier tasks can be effectively used for scoring candidate solutions of harder tasks and hence facilitating easy-to-hard generalization over different levels of tasks. Based on this insight, we propose a novel approach to scalable alignment, which firstly trains the (process-supervised) reward models on easy problems (e.g., level 1-3), and then uses them to evaluate the performance of policy models on hard problems. We show that such easy-to-hard generalization from evaluators can enable easy-to-hard generalizations in generators either through re-ranking or reinforcement learning (RL). Notably, our process-supervised 7b RL model and 34b model (reranking@1024) achieves an accuracy of 34.0% and 52.5% on MATH500, respectively, despite only using human supervision on easy problems. Our approach suggests a promising path toward AI systems that advance beyond the frontier of human supervision. * Equal contributions as leading authors. † Equal contributions as senior authors. 38th Conference on Neural Information Processing Systems (NeurIPS 2024).

For agents

The same record, over MCP at https://syntology.ai/mcp:

get_harvested_code_for_paper("2403.09472")
get_code_for_paper("2403.09472")
have("2403.09472")

The run record, dated, one paper per request, free:

curl https://syntology.ai/api/ran/2403.09472.json

A badge for a README (the split and the date, never a ratio):

[![Syntology run record](https://syntology.ai/api/ran/2403.09472.svg)](https://syntology.ai/paper/2403.09472)

Connect an agent — have() is free.