Abstract-CoT lets models reason with short discrete latent token sequences from a reserved vocabulary, using warm-up training and RL to match verbal CoT performance with up to 11.6x fewer tokens.
hub
Length-Controlled AlpacaEval: A Simple Way to Debias Automatic Evaluators
28 Pith papers cite this work. Polarity classification is still indexing.
28
Pith papers citing it
abstract
LLM-based auto-annotators have become a key component of the LLM development process due to their cost-effectiveness and scalability compared to human-based evaluation. However, these auto-annotators can introduce biases that are hard to remove. Even simple, known confounders such as preference for longer outputs remain in existing automated evaluation metrics. We propose a simple regression analysis approach for controlling biases in auto-evaluations. As a real case study, we focus on reducing the length bias of AlpacaEval, a fast and affordable benchmark for instruction-tuned LLMs that uses LLMs to estimate response quality. Despite being highly correlated with human preferences, AlpacaEval is known to favor models that generate longer outputs. We introduce a length-controlled AlpacaEval that aims to answer the counterfactual question: "What would the preference be if the model's and baseline's output had the same length?" To achieve this, we first fit a generalized linear model to predict the biased auto-annotator's preferences based on the mediators we want to control for (length difference) and other relevant features. We then obtain length-controlled preferences by predicting preferences while conditioning the GLM with a zero difference in lengths. Length-controlling not only improves the robustness of the metric to manipulations in model verbosity, but we also find that it increases the Spearman correlation with LMSYS Chatbot Arena from 0.94 to 0.98.
hub tools
citation-role summary
dataset 1
citation-polarity summary
roles
dataset 1polarities
use dataset 1representative citing papers
Continuous adversarial training in the embedding space produces a robust generalization bound for linear transformers that decreases with perturbation radius, tied to singular values of the embedding matrix, and motivates a new regularizer that improves real LLM jailbreak robustness-utility tradeoff
SalesLLM provides an automatic evaluation framework for LLM sales dialogues that correlates 0.98 with human experts and shows top models approaching human performance while weaker ones lag.
TiCo enables spoken dialogue models to follow explicit time constraints in generated responses using Spoken Time Markers and reinforcement learning with verifiable rewards, cutting duration error by 2.7x over its backbone.
DeepSeek-V2 delivers top-tier open-source LLM performance using only 21B active parameters by compressing the KV cache 93.3% and cutting training costs 42.5% via MLA and DeepSeekMoE.
RAG-Pref is a training-free RAG-based alignment technique that conditions LLMs on contrastive preference samples during inference, yielding over 3.7x average improvement in agentic attack refusals when combined with offline methods across five LLMs.
G-Zero uses the Hint-δ intrinsic reward to drive co-evolution between a Proposer and Generator via GRPO and DPO, providing a theoretical suboptimality guarantee for self-improvement from internal dynamics alone.
Bias-corrected LLM-as-a-Judge estimators can reverse true model orderings under shared calibration, and the paper supplies judge quality J and cross-model instability ΔJ as practical diagnostics for when such estimates are unreliable.
SKOP uses key-orthogonal projections to steer LLM activations while preserving attention patterns on focus tokens, cutting utility degradation by 5-7x and retaining over 95% of standard steering efficacy.
LLM safety judges flip verdicts on equivalent policy rewrites up to 9.1% of the time and cannot distinguish meaningful from meaningless changes, requiring new invariance-based reliability metrics.
DEPO uses historical data to build a data-dependent uncertainty bonus for exploration in online RLHF, yielding an adaptive regret bound and stronger empirical performance than baselines.
LocalAlign generates near-target adversarial examples via prompting and applies margin-aware alignment training to enforce tighter boundaries against prompt injection attacks.
TwinGate deploys a stateful dual-encoder system with asymmetric contrastive learning to detect decompositional jailbreaks in untraceable LLM traffic at high recall and low false-positive rate with negligible latency.
MultEval supports collaborative creation of LLM-as-a-judge criteria by surfacing disagreements via consensus-building methods, allowing iterative revisions with examples and history, and keeping transparent how human judgments become automated rules.
Analysis of the LMArena dataset reveals heavy topic skew and varying model rankings, leading to an interactive visualization tool for users to define custom evaluation priorities on LLM leaderboards.
Hybrid Policy Distillation unifies existing knowledge distillation methods for LLMs into a reweighted log-likelihood objective and introduces a hybrid forward-reverse KL approach with mixed data sampling to improve stability, efficiency, and performance.
S2H-DPO generates hierarchical prompt-driven preference pairs to improve multi-image reasoning in VLMs while keeping single-image performance intact.
BAR trains independent domain experts via separate mid-training, SFT, and RL pipelines then composes them with a MoE router to match monolithic retraining performance at lower cost and without catastrophic forgetting.
A unified incentive-score decomposition of preference optimization reveals the disentanglement band condition and reward calibration method that enables suppressing losers while preserving winners in LLM training.
AI models exhibit identity-contingent withholding, providing better clinical guidance on benzodiazepine tapering to physicians than laypeople in identical scenarios, with a measured decoupling gap of +0.38 and 13.1 percentage point drop in safety-critical action hit rates.
Relative density ratio optimization stabilizes direct density ratio estimation for language model alignment while preserving statistical consistency without assuming a Bradley-Terry preference model.
Embedding disruption re-triggers LLM internal safeguards to detect jailbreak prompts more effectively than standalone defenses.
The paper releases a 1,554-prompt consensus-labeled bank separating executable malicious code requests from security knowledge requests, validated by five-model majority labeling with Fleiss' kappa of 0.876.
LLM judges for human-AI coding co-creation show moderate performance (ROC-AUC 0.59) and low agreement, with co-creation success concentrating early in interactions.
citing papers explorer
-
Thinking Without Words: Efficient Latent Reasoning with Abstract Chain-of-Thought
Abstract-CoT lets models reason with short discrete latent token sequences from a reserved vocabulary, using warm-up training and RL to match verbal CoT performance with up to 11.6x fewer tokens.
-
Understanding and Improving Continuous Adversarial Training for LLMs via In-context Learning Theory
Continuous adversarial training in the embedding space produces a robust generalization bound for linear transformers that decreases with perturbation radius, tied to singular values of the embedding matrix, and motivates a new regularizer that improves real LLM jailbreak robustness-utility tradeoff
-
Sell More, Play Less: Benchmarking LLM Realistic Selling Skill
SalesLLM provides an automatic evaluation framework for LLM sales dialogues that correlates 0.98 with human experts and shows top models approaching human performance while weaker ones lag.
-
TiCo: Time-Controllable Spoken Dialogue Model
TiCo enables spoken dialogue models to follow explicit time constraints in generated responses using Spoken Time Markers and reinforcement learning with verifiable rewards, cutting duration error by 2.7x over its backbone.
-
DeepSeek-V2: A Strong, Economical, and Efficient Mixture-of-Experts Language Model
DeepSeek-V2 delivers top-tier open-source LLM performance using only 21B active parameters by compressing the KV cache 93.3% and cutting training costs 42.5% via MLA and DeepSeekMoE.
-
Leveraging RAG for Training-Free Alignment of LLMs
RAG-Pref is a training-free RAG-based alignment technique that conditions LLMs on contrastive preference samples during inference, yielding over 3.7x average improvement in agentic attack refusals when combined with offline methods across five LLMs.
-
G-Zero: Self-Play for Open-Ended Generation from Zero Data
G-Zero uses the Hint-δ intrinsic reward to drive co-evolution between a Proposer and Generator via GRPO and DPO, providing a theoretical suboptimality guarantee for self-improvement from internal dynamics alone.
-
Bias and Uncertainty in LLM-as-a-Judge Estimation
Bias-corrected LLM-as-a-Judge estimators can reverse true model orderings under shared calibration, and the paper supplies judge quality J and cross-model instability ΔJ as practical diagnostics for when such estimates are unreliable.
-
Don't Lose Focus: Activation Steering via Key-Orthogonal Projections
SKOP uses key-orthogonal projections to steer LLM activations while preserving attention patterns on focus tokens, cutting utility degradation by 5-7x and retaining over 95% of standard steering efficacy.
-
Beyond Accuracy: Policy Invariance as a Reliability Test for LLM Safety Judges
LLM safety judges flip verdicts on equivalent policy rewrites up to 9.1% of the time and cannot distinguish meaningful from meaningless changes, requiring new invariance-based reliability metrics.
-
Data-dependent Exploration for Online Reinforcement Learning from Human Feedback
DEPO uses historical data to build a data-dependent uncertainty bonus for exploration in online RLHF, yielding an adaptive regret bound and stronger empirical performance than baselines.
-
LocalAlign: Enabling Generalizable Prompt Injection Defense via Generation of Near-Target Adversarial Examples for Alignment Training
LocalAlign generates near-target adversarial examples via prompting and applies margin-aware alignment training to enforce tighter boundaries against prompt injection attacks.
-
TwinGate: Stateful Defense against Decompositional Jailbreaks in Untraceable Traffic via Asymmetric Contrastive Learning
TwinGate deploys a stateful dual-encoder system with asymmetric contrastive learning to detect decompositional jailbreaks in untraceable LLM traffic at high recall and low false-positive rate with negligible latency.
-
MultEval: Supporting Collaborative Alignment for LLM-as-a-Judge Evaluation Criteria
MultEval supports collaborative creation of LLM-as-a-judge criteria by surfacing disagreements via consensus-building methods, allowing iterative revisions with examples and history, and keeping transparent how human judgments become automated rules.
-
Who Defines "Best"? Towards Interactive, User-Defined Evaluation of LLM Leaderboards
Analysis of the LMArena dataset reveals heavy topic skew and varying model rankings, leading to an interactive visualization tool for users to define custom evaluation priorities on LLM leaderboards.
-
Hybrid Policy Distillation for LLMs
Hybrid Policy Distillation unifies existing knowledge distillation methods for LLMs into a reweighted log-likelihood objective and introduces a hybrid forward-reverse KL approach with mixed data sampling to improve stability, efficiency, and performance.
-
S2H-DPO: Hardness-Aware Preference Optimization for Vision-Language Models
S2H-DPO generates hierarchical prompt-driven preference pairs to improve multi-image reasoning in VLMs while keeping single-image performance intact.
-
Train Separately, Merge Together: Modular Post-Training with Mixture-of-Experts
BAR trains independent domain experts via separate mid-training, SFT, and RL pipelines then composes them with a MoE router to match monolithic retraining performance at lower cost and without catastrophic forgetting.
-
Towards Disentangled Preference Optimization Dynamics: Suppress the Loser, Preserve the Winner
A unified incentive-score decomposition of preference optimization reveals the disentanglement band condition and reward calibration method that enables suppressing losers while preserving winners in LLM training.
-
IatroBench: Pre-Registered Evidence of Iatrogenic Harm from AI Safety Measures
AI models exhibit identity-contingent withholding, providing better clinical guidance on benzodiazepine tapering to physicians than laypeople in identical scenarios, with a measured decoupling gap of +0.38 and 13.1 percentage point drop in safety-critical action hit rates.
-
Relative Density Ratio Optimization for Stable and Statistically Consistent Model Alignment
Relative density ratio optimization stabilizes direct density ratio estimation for language model alignment while preserving statistical consistency without assuming a Bradley-Terry preference model.
-
Re-Triggering Safeguards within LLMs for Jailbreak Detection
Embedding disruption re-triggers LLM internal safeguards to detect jailbreak prompts more effectively than standalone defenses.
-
A Validated Prompt Bank for Malicious Code Generation: Separating Executable Weapons from Security Knowledge in 1,554 Consensus-Labeled Prompts
The paper releases a 1,554-prompt consensus-labeled bank separating executable malicious code requests from security knowledge requests, validated by five-model majority labeling with Fleiss' kappa of 0.876.
-
LLM-as-a-Judge for Human-AI Co-Creation: A Reliability-Aware Evaluation Framework for Coding
LLM judges for human-AI coding co-creation show moderate performance (ROC-AUC 0.59) and low agreement, with co-creation success concentrating early in interactions.
-
Judging the Judges: A Systematic Evaluation of Bias Mitigation Strategies in LLM-as-a-Judge Pipelines
Style bias dominates LLM-as-a-Judge systems far more than position bias, with debiasing strategies providing model-dependent gains and public tools released for replication.
-
MedConclusion: A Benchmark for Biomedical Conclusion Generation from Structured Abstracts
MedConclusion is a 5.7M-instance benchmark dataset for generating biomedical conclusions from structured PubMed abstracts, with LLM evaluations showing conclusion writing differs from summarization and that judge choice affects scores.
-
LLMs-as-Judges: A Comprehensive Survey on LLM-based Evaluation Methods
A survey that organizes LLMs-as-judges research into functionality, methodology, applications, meta-evaluation, and limitations.
- CyclicJudge: Mitigating Judge Bias Efficiently in LLM-based Evaluation