SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2603.14889 · 2026

SDiaReward: Modeling and Benchmarking Spoken Dialogue Rewards with Modality and Colloquialness

Zhou Zhao, Yuhan Wang, Xize Cheng, Yifu Chen, Changhao Pan, Jingyu Lu, Fan Zhuo, Xueyi Pu, Chenyuhao Wen, Tianle Liang

arXiv · PDF · Open in the Atlas

Code that ran

11 of 15 samples (functions and classes) lifted from this paper's linked repositories executed on a synthesized input in Syntology's jail (1 of them constructed an object rather than computing a result; 0 honoured the contract we drafted, 1 violated it and 3 have no contract verdict; 7 failures of our instrument; 4 no recorded run), measured 2026-11-22; 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/2603.14889.json · badge: /api/ran/2603.14889.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
Anjok07/ultimatevocalremovergui canonical (classifier label)
classifier basis: census
7 ran · 3 no recorded run
MM-Speech/SDiaReward community
Papers with Code archive
4 ran · 1 no recorded run
FunctionStatusWhere it lives
AttentionPooling Ran
ran on a synthesized input · constructed an object; no result was computed
MM-Speech/SDiaReward/model/modeling_qwen_omni_thinker_reward.py
code served (permissive licence) · get_code("56d866d35d4af648")
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, through the batch API) and returned an instance of its own class, AttentionPooling: the run constructed the object and did not call it, so no result was computed. 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
import torch.nn as nn

torch.manual_seed(0)

class AttentionPooling(nn.Module):
    """Attention pooling for sentence-level representation"""

    def __init__(self, hidden_size: int):
        super().__init__()
        self.attn = nn.Linear(hidden_size, 1)

    def forward(self, hidden_states, attention_mask):
        attn_scores = self.attn(hidden_states).squeeze(-1)
        attn_scores = attn_scores.masked_fill(attention_mask == 0, -1000000000.0)
        attn_weights = torch.softmax(attn_scores, dim=-1)
        pooled = torch.sum(hidden_states * attn_weights.unsqueeze(-1), dim=1)
        return (pooled, attn_weights)

# Setup: create instance and forward pass arguments
hidden_size = 4
model = AttentionPooling(hidden_size=hidden_size)

# Create batch of hidden states: (batch_size=2, seq_length=3, hidden_size=4)
hidden_states = torch.randn(2, 3, hidden_size)

# Create attention mask: (batch_size=2, seq_length=3)
# 1 means attend, 0 means mask out
attention_mask = torch.ones(2, 3, dtype=torch.long)
attention_mask[0, 2] = 0  # mask out last token in first example

ARGS = {
    "hidden_size": hidden_size,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type AttentionPooling — the run constructed an instance of its own class and did not call it; no shape is stated — the instrument coerces any output to an array; for a AttentionPooling 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
LastTokenPooling Ran
ran on a synthesized input
MM-Speech/SDiaReward/model/modeling_qwen_omni_thinker_reward.py
code served (permissive licence) · get_code("05e376e6778ec9f2")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, class construct-then-call prompt, model not recorded) and returned tuple. 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
import torch.nn as nn

class LastTokenPooling(nn.Module):

    def forward(self, hidden_states: torch.Tensor, mask: torch.Tensor):
        idx = mask.long().sum(dim=1) - 1
        idx = idx.clamp(min=0)
        b = torch.arange(hidden_states.size(0), device=hidden_states.device)
        return (hidden_states[b, idx], None)


INIT_ARGS = {}

ARGS = {
    'hidden_states': torch.randn(2, 5, 4),
    'mask': torch.tensor([[1, 1, 1, 0, 0], [1, 1, 1, 1, 0]], dtype=torch.float32)
}

if __name__ == '__main__':
    model = LastTokenPooling(**INIT_ARGS)
    output = model(**ARGS)
    print("Output:", output)
    print("Pooled shape:", output[0].shape)
Setup block drafted by
a language model, class construct-then-call prompt, model not recorded
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: torch and numpy seeded to 0 by the harness immediately before the class was constructed from INIT_ARGS (fixture_check.py)
MeanPooling Ran
ran on a synthesized input
MM-Speech/SDiaReward/model/modeling_qwen_omni_thinker_reward.py
code served (permissive licence) · get_code("ac1c99e669b0bd5f")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, class construct-then-call prompt, model not recorded) and returned tuple. 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
import torch.nn as nn

class MeanPooling(nn.Module):

    def forward(self, hidden_states: torch.Tensor, mask: torch.Tensor):
        w = mask.unsqueeze(-1).to(hidden_states.dtype)
        denom = w.sum(dim=1).clamp(min=1.0)
        pooled = (hidden_states * w).sum(dim=1) / denom
        return (pooled, None)


INIT_ARGS = {}

ARGS = {
    'hidden_states': torch.randn(2, 4, 8),
    'mask': torch.ones(2, 4, dtype=torch.long)
}

if __name__ == '__main__':
    model = MeanPooling(**INIT_ARGS)
    output = model(**ARGS)
    print("Output shape:", output[0].shape)
    print("Output:", output)
Setup block drafted by
a language model, class construct-then-call prompt, model not recorded
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: torch and numpy seeded to 0 by the harness immediately before the class was constructed from INIT_ARGS (fixture_check.py)
atan2 Ran
ran; our contract draft was wrong, not the code
Anjok07/ultimatevocalremovergui/demucs/filtering.py
code served (permissive licence) · get_code("9e36f722330708b6")
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, through the batch API) and returned Tensor, ndim 2, dtype family float, shape (2, 4) 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

torch.manual_seed(0)

# Create two small tensors with the same shape
y = torch.randn(2, 4)
x = torch.randn(2, 4)

ARGS = {
    "y": y,
    "x": x,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
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 float, shape (2, 4) — shape (2, 4) 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
build_pooler Ran
ran; our contract draft was wrong, not the code
MM-Speech/SDiaReward/model/modeling_qwen_omni_thinker_reward.py
code served (permissive licence) · get_code("259e29166466af50")
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, through the batch API) and returned AttentionPooling. 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
import torch.nn as nn

torch.manual_seed(0)

# Test with "attention" pooling type
pooling_type = "attention"
hidden_size = 4

ARGS = {
    "pooling_type": pooling_type,
    "hidden_size": hidden_size,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type AttentionPooling — a AttentionPooling output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a AttentionPooling 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
chose_norm Ran
ran; our contract draft was wrong, not the code
Anjok07/ultimatevocalremovergui/demucs/tasnet.py
code served (permissive licence) · get_code("c20ada1396105c22")
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, through the batch API) and returned GlobalLayerNorm. 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
import torch.nn as nn

torch.manual_seed(0)

ARGS = {
    'norm_type': 'gLN',
    'channel_size': 16
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type GlobalLayerNorm — a GlobalLayerNorm output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a GlobalLayerNorm 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
downsample Ran
ran; the contract check could not run (RAISES, DEP_MISSING or TIMEOUT)
Anjok07/ultimatevocalremovergui/demucs/model.py
code served (permissive licence) · get_code("07fa2ca4b8bf0830")
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, through the batch API) and returned Tensor, ndim 3, dtype family float, shape (2, 4, 3) 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

torch.manual_seed(0)

# Create a small 3D tensor with shape (batch=2, channels=4, time=9)
# The function indexes with ::stride, so we need enough time steps
x = torch.randn(2, 4, 9)

# Create a stride parameter
stride = 3

ARGS = {
    "x": x,
    "stride": stride,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
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, 4, 3) — shape (2, 4, 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
gather_sources Ran
ran, violated the contract we drafted
Anjok07/ultimatevocalremovergui/separate.py
code served (permissive licence) · get_code("e870d9e2657f0bb3")
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, through the batch API) and returned tuple, ndim 1, dtype family float; no output shape is stated. The image it ran in was not recorded. If your copy returns the same type and ndim 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)

primary_stem_name = "model_backbone_encoder"
secondary_stem_name = "feature_extractor_decoder"
secondary_sources = {
    "backbone": True,
    "encoder": False,
    "decoder": True,
    "extractor": False
}

ARGS = {
    "primary_stem_name": primary_stem_name,
    "secondary_stem_name": secondary_stem_name,
    "secondary_sources": secondary_sources
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type tuple, ndim 1, dtype family float — a tuple output has no array shape of its own; none is stated
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
overlap_and_add Ran
ran; our contract draft was wrong, not the code
Anjok07/ultimatevocalremovergui/demucs/tasnet.py
code served (permissive licence) · get_code("6fc6cdb0729e7905")
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, through the batch API) and returned Tensor, ndim 2, dtype family float, shape (2, 16) 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 math

torch.manual_seed(0)

# The function performs overlap-and-add on signal frames
# signal: tensor with shape (*outer_dimensions, frames, frame_length)
# frame_step: integer step size between frames

# Create a small signal tensor
# Shape: (2, 3, 8) meaning batch_size=2, frames=3, frame_length=8
signal = torch.randn(2, 3, 8)

# frame_step should be less than frame_length for overlap-and-add to make sense
# Using frame_step=4 with frame_length=8 creates 50% overlap
frame_step = 4

ARGS = {
    "signal": signal,
    "frame_step": frame_step,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
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 float, shape (2, 16) — shape (2, 16) 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
pad1d Ran
ran; the contract check could not run (RAISES, DEP_MISSING or TIMEOUT)
Anjok07/ultimatevocalremovergui/demucs/hdemucs.py
code served (permissive licence) · get_code("3bc760123c304a99")
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, through the batch API) and returned Tensor, ndim 2, dtype family float, shape (2, 7) 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 typing as tp

torch.manual_seed(0)

# Create a small tensor with shape (batch=2, dim=4)
x = torch.randn(2, 4)

# Tuple of (padding_left, padding_right)
paddings = (1, 2)

# Mode for padding
mode = 'reflect'

# Value for constant padding
value = 0.0

ARGS = {
    'x': x,
    'paddings': paddings,
    'mode': mode,
    'value': value
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
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 float, shape (2, 7) — shape (2, 7) 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
upsample Ran
ran; our contract draft was wrong, not the code
Anjok07/ultimatevocalremovergui/demucs/model.py
code served (permissive licence) · get_code("5cf51541da88d14a")
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, through the batch API) and returned Tensor, ndim 3, dtype family float, shape (2, 4, 4) 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 as th

th.manual_seed(0)

# Create a small 3D tensor (batch, channels, time)
# batch=2, channels=4, time=3
x = th.randn(2, 4, 3)

# stride parameter - upsampling factor
stride = 2

ARGS = {
    "x": x,
    "stride": stride,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
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, 4, 4) — shape (2, 4, 4) 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
QwenOmniThinkerReward Not yet run
unverified (harvested, no recorded run)
MM-Speech/SDiaReward/model/modeling_qwen_omni_thinker_reward.py
code served (permissive licence) · get_code("10092f16f165a8c4")
How it ran

no invocation recorded.

expectation_maximization Not yet run
unverified (harvested, no recorded run)
Anjok07/ultimatevocalremovergui/demucs/filtering.py
code served (permissive licence) · get_code("70d7a8250b945826")
How it ran

no invocation recorded.

tensor_chunk Not yet run
unverified (harvested, no recorded run)
Anjok07/ultimatevocalremovergui/demucs/apply.py
code served (permissive licence) · get_code("c1164f5f84afdafa")
How it ran

no invocation recorded.

wiener Not yet run
unverified (harvested, no recorded run)
Anjok07/ultimatevocalremovergui/demucs/filtering.py
code served (permissive licence) · get_code("54c7cf43e5e2bb89")
How it ran

no invocation recorded.

Also present in other papers' code

Of 15 samples lifted from this paper's linked repositories, 15 carry a function identity and 0 do not. Of the 15 distinct functions with an identity, 9 are also present in other papers whose code we hold (6 with at least one paper dated earlier, 3 the same month, 0 later) and 6 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
chose_norm 2 other papers
2 dated earlier
1809.07454 (2018-09)
Conv-TasNet: Surpassing Ideal Time-Frequency Magnitude Masking for Speech Separation
2211.08553 (2022-11)
Hybrid Transformers for Music Source Separation
downsample 2 other papers
2 dated earlier
1809.07454 (2018-09)
Conv-TasNet: Surpassing Ideal Time-Frequency Magnitude Masking for Speech Separation
2211.08553 (2022-11)
Hybrid Transformers for Music Source Separation
overlap_and_add 2 other papers
2 dated earlier
1809.07454 (2018-09)
Conv-TasNet: Surpassing Ideal Time-Frequency Magnitude Masking for Speech Separation
2211.08553 (2022-11)
Hybrid Transformers for Music Source Separation
upsample 2 other papers
2 dated earlier
1809.07454 (2018-09)
Conv-TasNet: Surpassing Ideal Time-Frequency Magnitude Masking for Speech Separation
2211.08553 (2022-11)
Hybrid Transformers for Music Source Separation
atan2 1 other paper
1 the same month
2603.09187 (2026-03)
The Costs of Reproducibility in Music Separation Research: a Replication of Band-Split RNN
expectation_maximization 1 other paper
1 the same month
2603.09187 (2026-03)
The Costs of Reproducibility in Music Separation Research: a Replication of Band-Split RNN
pad1d 1 other paper
1 dated earlier
2601.15348 (2026-01)
Abusive music transformation using GenAI and LLMs
tensor_chunk 1 other paper
1 dated earlier
2601.15348 (2026-01)
Abusive music transformation using GenAI and LLMs
wiener 1 other paper
1 the same month
2603.09187 (2026-03)
The Costs of Reproducibility in Music Separation Research: a Replication of Band-Split RNN

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

The rapid evolution of end-to-end spoken dialogue systems demands transcending mere textual semantics to incorporate paralinguistic nuances and the spontaneous nature of human conversation. However, current methods struggle with two critical gaps: the modality gap, involving prosody and emotion, and the colloquialness gap, distinguishing written scripts from natural speech. To address these challenges, we introduce SDiaReward, an end-to-end multiturn reward model trained on SDiaReward-Dataset, a novel collection of episode-level preference pairs explicitly targeting these gaps. It operates directly on full multi-turn speech episodes and is optimized with pairwise preference supervision, enabling joint assessment of modality and colloquialness in a single evaluator. We further establish ESDR-Bench, a stratified benchmark for robust episode-level evaluation. Experiments demonstrate that SDi-aReward achieves state-of-the-art pairwise preference accuracy, significantly outperforming general-purpose audio LLMs. Further analysis suggests that SDiaReward captures relative conversational expressiveness beyond superficial synthesis cues, improving generalization across domains and recording conditions. Code, data, and demos are available at https: //github.com/MM-Speech/SDiaReward/.

For agents

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

get_harvested_code_for_paper("2603.14889")
get_code_for_paper("2603.14889")
have("2603.14889")

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

curl https://syntology.ai/api/ran/2603.14889.json

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

[![Syntology run record](https://syntology.ai/api/ran/2603.14889.svg)](https://syntology.ai/paper/2603.14889)

Connect an agent — have() is free.