菜单
Token-adaptive preference optimization framework using min-max formulation to reduce multimodal LLM hallucination. Achieves 50% hallucination reduction using min-max distributional robustness while preserving visual grounding.
默认使用会先检查来源的 Prompt；你也可以切换为直接命令，或下载本地副本。
决定是否安装前，请先阅读 SKILL.md，以及 SkillsMP 当前展示的配套文件。
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
直接命令不会经过审查 Prompt；运行前请先检查来源。
npx skills add https://github.com/ADu2021/skillXiv --skill tars-token-adaptive-hallucination-reduction命令会保持在同一行。复制前请横向滚动并检查完整内容。
想先保存到本地？可下载 SkillsMP 当前能够提供的文件。
基于 SOC 职业分类
正在显示 SKILL.md
| name | tars-token-adaptive-hallucination-reduction |
| title | TARS Min-Max Token-Adaptive Preference Strategy |
| version | 0.0.2 |
| engine | skillxiv-v0.0.2-claude-opus-4.6 |
| license | MIT |
| url | https://arxiv.org/abs/2507.21584 |
| keywords | ["multimodal-llm","hallucination-reduction","preference-optimization","token-adaptive","robustness"] |
| description | Token-adaptive preference optimization framework using min-max formulation to reduce multimodal LLM hallucination. Achieves 50% hallucination reduction using min-max distributional robustness while preserving visual grounding. |
TARS addresses the critical problem of hallucination in multimodal large language models through a novel token-level adaptive preference optimization approach. Rather than treating preference signals as fixed targets, TARS uses min-max optimization to train models that are robust to preference label uncertainty while maintaining visual grounding.
The key insight is that hallucinations arise when language models overfit to spurious patterns in preference data, disconnecting outputs from actual visual content. TARS reformulates preference optimization as a min-max problem:
This approach reduces hallucinations from 26.4% to 13.2% using only 4.8k preference samples, achieving state-of-the-art performance.
The framework consists of:
Step 1: Encode preferences as token-level distributions
Convert pairwise preference labels into token-level target distributions:
import torch
import torch.nn.functional as F
from typing import List, Tuple, Dict
class PreferenceEncoder:
"""Encodes human preferences as token-level distributions"""
def __init__():
.tokenizer = tokenizer
vocab_size = vocabulary_size
() -> [torch.Tensor, torch.Tensor]:
preferred_tokens = .tokenizer.encode(preferred_text)
dispreferred_tokens = .tokenizer.encode(dispreferred_text)
max_len = ((preferred_tokens), (dispreferred_tokens))
preferred_dist = torch.zeros(max_len, .vocab_size)
dispreferred_dist = torch.zeros(max_len, .vocab_size)
i, token_id (preferred_tokens):
preferred_dist[i, token_id] =
i, token_id (dispreferred_tokens):
dispreferred_dist[i, token_id] =
preferred_dist, dispreferred_dist
() -> torch.Tensor:
all_prefs = []
preferred, dispreferred pairs:
pref_dist, _ = .encode_preference_pair(preferred, dispreferred)
all_prefs.append(pref_dist)
max_len = (p.shape[] p all_prefs)
batch_prefs = torch.zeros((all_prefs), max_len, .vocab_size)
i, pref (all_prefs):
batch_prefs[i, :pref.shape[], :] = pref
batch_prefs
This represents preferences as target probability distributions over tokens.
Step 2: Implement min-max optimization with perturbation
The core training loop applies min-max optimization to achieve robustness:
class MinMaxPreferenceOptimizer:
"""Min-max optimization for robust preference learning"""
def __init__(self, model, tokenizer, epsilon: float = 0.1):
self.model = model
self.tokenizer = tokenizer
self.epsilon = epsilon # Perturbation budget
self.vocab_size = model.config.vocab_size
def compute_preference_loss(self, logits: torch.Tensor,
target_dist: torch.Tensor,
positions: torch.Tensor) -> torch.Tensor:
"""
Compute loss between predicted logits and preference targets.
Args:
logits: Model output logits, shape (batch, seq_len, vocab_size)
target_dist: Target probability distribution over tokens
positions: Which positions correspond to preference targets
Returns:
Scalar loss value
"""
# Extract logits at preference positions
pred_probs = F.softmax(logits, dim=-1)
# KL divergence from model distribution to target
loss = F.kl_div(
F.log_softmax(logits, dim=-1),
target_dist.clamp(min=1e-8),
reduction='batchmean'
)
return loss
def perturbation_step(self, target_dist: torch.Tensor,
step_size: float = 0.01) -> torch.Tensor:
"""
Inner optimization: find worst-case perturbation within semantic budget.
Args:
target_dist: Current preference target
step_size: Optimization step size
Returns:
Perturbed distribution that remains semantically valid
"""
perturbed = target_dist.clone().detach().requires_grad_(True)
optimizer = torch.optim.Adam([perturbed], lr=step_size)
for _ in range(5): # Inner loop iterations
# Compute model loss on perturbed targets
logits = self.model(...) # Forward pass
loss = self.compute_preference_loss(logits, perturbed, mask=None)
# Maximize loss (adversarial): find worst perturbation
loss_adv = -loss
optimizer.zero_grad()
loss_adv.backward()
optimizer.step()
# Project back to semantic constraint set
perturbed.data = self._project_to_semantic_budget(perturbed.data)
return perturbed.detach()
def _project_to_semantic_budget(self, dist: torch.Tensor) -> torch.Tensor:
"""
Project distribution back to semantic constraint set.
Keep similar tokens with higher weight.
"""
# Reproject to valid probability distribution
dist = F.softmax(dist, dim=-1)
# Constraint: perturbed distribution cannot exceed epsilon away from original
# (in terms of TV divergence or other semantic distance metric)
# Simplified: clamp large changes
return dist
def min_max_training_step(self, images: torch.Tensor,
prompts: torch.Tensor,
preference_targets: torch.Tensor) -> Tuple[torch.Tensor, float]:
"""
Single min-max training step: maximize robustness to preference perturbations.
Args:
images: Visual input
prompts: Text prompts
preference_targets: Target distributions for next tokens
Returns:
(updated_model_state, loss_value)
"""
# Outer loop: optimize model parameters
optimizer = torch.optim.AdamW(self.model.parameters(), lr=1e-5)
for outer_step in range(3): # Outer iterations
# Inner loop: find worst perturbation
perturbed_targets = self.perturbation_step(preference_targets)
# Forward pass with model
outputs = self.model(images, prompts)
logits = outputs.logits
# Compute loss on perturbed targets
loss = self.compute_preference_loss(logits, perturbed_targets, positions=None)
# Outer optimization: minimize loss on worst-case perturbation
optimizer.zero_grad()
loss.backward()
optimizer.step()
return self.model.state_dict(), loss.item()
This implements the core min-max algorithm that creates robustness to preference uncertainty.
Step 3: Implement token-level adaptive steering
Apply different preference strengths to different tokens:
class TokenAdaptivePreference:
"""Adaptively applies preference steering per token"""
def __init__(self, model):
self.model = model
def compute_token_importance(self, logits: torch.Tensor,
target_dist: torch.Tensor) -> torch.Tensor:
"""
Compute importance weight for each token based on preference strength.
Args:
logits: Model predictions
target_dist: Preference targets
Returns:
Importance weights of shape (batch, seq_len)
"""
pred_probs = F.softmax(logits, dim=-1)
# Importance = entropy difference (high-entropy targets need stronger steering)
target_entropy = -(target_dist * torch.log(target_dist + 1e-8)).sum(dim=-1)
pred_entropy = -(pred_probs * torch.log(pred_probs + 1e-8)).sum(dim=-1)
importance = (target_entropy - pred_entropy).clamp(min=0)
# Normalize to [0, 1]
importance = importance / (importance.max() + 1e-8)
return importance
def apply_adaptive_preference(self, logits: torch.Tensor,
target_dist: torch.Tensor,
base_strength: float = 1.0) -> torch.Tensor:
"""
Shift logits toward preference targets with token-adaptive strength.
Args:
logits: Model logits
target_dist: Preference targets
base_strength: Base preference steering strength
Returns:
Modified logits
"""
# Compute per-token importance weights
importance = self.compute_token_importance(logits, target_dist)
# Compute preference direction
pred_probs = F.softmax(logits, dim=-1)
pref_direction = target_dist - pred_probs # Shape: (batch, seq_len, vocab)
# Apply adaptive steering
batch_size, seq_len, vocab_size = logits.shape
for b in range(batch_size):
for t in range(seq_len):
# Scale steering by token importance
strength = base_strength * importance[b, t].item()
# Move logits toward preference targets
logits[b, t, :] = logits[b, t, :] + strength * pref_direction[b, t, :]
return logits
This enables the system to focus preference steering on critical tokens.
Step 4: Preserve visual grounding during steering
Ensure that preference steering doesn't disconnect from image content:
class VisualGroundingPreserver:
"""Ensures preference optimization preserves visual grounding"""
def __init__(self, model, vision_encoder):
self.model = model
self.vision_encoder = vision_encoder
def compute_visual_consistency(self, image: torch.Tensor,
output_tokens: torch.Tensor) -> float:
"""
Measure how well output is grounded in visual content.
Compute correlation between image features and output token embeddings.
"""
# Extract visual features
image_features = self.vision_encoder(image) # (image_dim,)
# Extract output token embeddings
token_embeddings = self.model.get_token_embeddings(output_tokens) # (seq_len, emb_dim)
# Compute attention-weighted consistency
batch_size, seq_len, vocab_size = output_tokens.shape
similarities = []
for t in range(seq_len):
# Similarity between this token's embedding and image features
token_emb = token_embeddings[t]
similarity = F.cosine_similarity(
image_features.unsqueeze(0),
token_emb.unsqueeze(0)
)
similarities.append(similarity)
# Average similarity across tokens
visual_consistency = torch.mean(torch.stack(similarities))
return visual_consistency.item()
def grounding_aware_loss(self, logits: torch.Tensor,
target_dist: torch.Tensor,
image: torch.Tensor,
lambda_grounding: float = 0.1) -> torch.Tensor:
"""
Combine preference loss with visual grounding preservation.
Args:
logits: Model predictions
target_dist: Preference targets
image: Visual input
lambda_grounding: Weight for grounding constraint
Returns:
Total loss that preserves grounding
"""
# Preference alignment loss
preference_loss = F.kl_div(
F.log_softmax(logits, dim=-1),
target_dist,
reduction='batchmean'
)
# Visual grounding penalty
predicted_tokens = torch.argmax(logits, dim=-1)
grounding_score = self.compute_visual_consistency(image, predicted_tokens)
# Total loss: preference optimization + grounding preservation
total_loss = preference_loss - lambda_grounding * grounding_score
return total_loss
This ensures hallucination reduction doesn't sacrifice visual fidelity.
Step 5: Training loop with all components
Integrate all components into an end-to-end training procedure:
class TARSTrainer:
"""Complete TARS training with preference, perturbation, and grounding"""
def __init__(self, model, vision_encoder, tokenizer,
epsilon: float = 0.1, lambda_grounding: float = 0.1):
self.model = model
self.vision_encoder = vision_encoder
self.pref_encoder = PreferenceEncoder(tokenizer, model.config.vocab_size)
self.optimizer = MinMaxPreferenceOptimizer(model, tokenizer, epsilon)
self.token_adaptive = TokenAdaptivePreference(model)
self.grounding = VisualGroundingPreserver(model, vision_encoder)
self.lambda_grounding = lambda_grounding
def train_step(self, batch: Dict) -> float:
"""
Single TARS training step.
Args:
batch: Contains 'images', 'prompts', 'preferred_text', 'dispreferred_text'
Returns:
Loss value
"""
images = batch['images']
prompts = batch['prompts']
preferred = batch['preferred_text']
dispreferred = batch['dispreferred_text']
# Encode preferences as distributions
pref_targets = self.pref_encoder.create_preference_labels(
list(zip(preferred, dispreferred))
)
# Forward pass
outputs = self.model(images, prompts)
logits = outputs.logits
# Apply token-adaptive preference steering
adapted_logits = self.token_adaptive.apply_adaptive_preference(
logits, pref_targets, base_strength=1.0
)
# Compute loss with visual grounding preservation
loss = self.grounding.grounding_aware_loss(
adapted_logits, pref_targets, images,
lambda_grounding=self.lambda_grounding
)
# Min-max optimization step
self.model.train()
loss.backward()
torch.nn.utils.clip_grad_norm_(self.model.parameters(), 1.0)
optimizer = torch.optim.AdamW(self.model.parameters(), lr=1e-5)
optimizer.step()
optimizer.zero_grad()
return loss.item()
def train_epoch(self, dataloader, num_epochs: int = 3):
"""Train for multiple epochs"""
for epoch in range(num_epochs):
total_loss = 0.0
for batch_idx, batch in enumerate(dataloader):
loss = self.train_step(batch)
total_loss += loss
if batch_idx % 100 == 0:
print(f"Epoch {epoch}, Batch {batch_idx}, Loss: {loss:.4f}")
avg_loss = total_loss / len(dataloader)
print(f"Epoch {epoch} average loss: {avg_loss:.4f}")
This trains models with robust preference optimization while preserving visual grounding.
When to use TARS:
When NOT to use TARS:
Key hyperparameters:
epsilon: Perturbation budget (0.05-0.15 typical; larger = more robust but slower)lambda_grounding: Grounding preservation weight (0.05-0.2; higher = more visual fidelity)base_strength: Preference steering magnitude (0.5-2.0)inner_steps: Perturbation loop iterations (3-7 typical)outer_steps: Model parameter update iterations (2-4 typical)Expected improvements:
Dataset requirements:
TARS: MinMax Token-Adaptive Preference Strategy for Hallucination Reduction. arXiv:2507.21584