Greedy random search recovers token sequences that elicit harmful response prefixes from LLMs without meaningful instructions, showing natural backdoors are present yet require more effort than semantic attacks.
Title resolution pending
15 Pith papers cite this work. Polarity classification is still indexing.
15
Pith papers citing it
citation-role summary
background 2
baseline 1
citation-polarity summary
representative citing papers
DuoAttention identifies retrieval heads requiring full KV cache and streaming heads using constant-length cache to reduce memory and latency in long-context LLM inference.
DEL is a new loss for LLM numerical learning that applies supervised digit entropy optimization and extends to floating-point numbers, showing improved accuracy and distance metrics over prior methods on math benchmarks.
ξ-DPO rewrites the preference objective as minimizing distance to optimal margins and defines reward as a chosen-to-rejected ratio, yielding a bounded, interpretable margin ξ set directly from the initial reward-gap distribution.
A single-pass black-box method models LLM outputs as dynamical systems via Koopman operators to detect hallucinations with claimed state-of-the-art accuracy and lower cost.
CleanBase identifies malicious documents in RAG databases by detecting cliques in a semantic similarity graph constructed using embedding models and a statistical threshold.
RewardBench 2 is a new benchmark that supplies challenging fresh human prompts for reward model evaluation, yielding lower average scores but higher correlation with downstream best-of-N sampling and RLHF training performance.
WildGuard is a new open moderation model and dataset for LLM safety that identifies harmful prompts, risky responses, and refusal rates, achieving SOTA open-source performance and sometimes exceeding GPT-4 while cutting jailbreak success from 79.8% to 2.4%.
Extending language model context length enables LMMs to process over 200K visual tokens from long videos without video training, achieving SOTA on Video-MME via dense frame sampling.
Gradual fine-tuning that removes explicit CoT steps lets GPT-2 Small reach 99% accuracy on 9x9 multiplication and Mistral 7B exceed 50% on GSM8K with no intermediate outputs.
DPOP is a new loss function that prevents DPO from lowering preferred response likelihoods and outperforms standard DPO on diverse datasets, MT-Bench, and enables Smaug-72B to exceed 80% on the Open LLM Leaderboard.
FragileFlow formalizes margin-aware error flow and applies spectral control through a calibrated margin buffer and class-wise risk matrix, supported by a PAC-Bayes bound, to enhance worst-class robustness in foundation model adaptation while preserving clean accuracy.
STELLA aligns ESM3 bimodal sequence-structure encodings with Llama-3.1-8B text modeling to claim state-of-the-art results on protein functional description prediction and enzyme-catalyzed reaction prediction.
A survey deriving a unified policy gradient framework for LLM post-training methods and providing technical comparisons of PPO, GRPO, DPO variants.
citing papers explorer
-
On the Hardness of Junking LLMs
Greedy random search recovers token sequences that elicit harmful response prefixes from LLMs without meaningful instructions, showing natural backdoors are present yet require more effort than semantic attacks.
-
DuoAttention: Efficient Long-Context LLM Inference with Retrieval and Streaming Heads
DuoAttention identifies retrieval heads requiring full KV cache and streaming heads using constant-length cache to reduce memory and latency in long-context LLM inference.
-
DEL: Digit Entropy Loss for Numerical Learning of Large Language Models
DEL is a new loss for LLM numerical learning that applies supervised digit entropy optimization and extends to floating-point numbers, showing improved accuracy and distance metrics over prior methods on math benchmarks.
-
$\xi$-DPO: Direct Preference Optimization via Ratio Reward Margin
ξ-DPO rewrites the preference objective as minimizing distance to optimal margins and defines reward as a chosen-to-rejected ratio, yielding a bounded, interpretable margin ξ set directly from the initial reward-gap distribution.
-
Low-Cost Black-Box Detection of LLM Hallucinations via Dynamical System Prediction
A single-pass black-box method models LLM outputs as dynamical systems via Koopman operators to detect hallucinations with claimed state-of-the-art accuracy and lower cost.
-
CleanBase: Detecting Malicious Documents in RAG Knowledge Databases
CleanBase identifies malicious documents in RAG databases by detecting cliques in a semantic similarity graph constructed using embedding models and a statistical threshold.
-
RewardBench 2: Advancing Reward Model Evaluation
RewardBench 2 is a new benchmark that supplies challenging fresh human prompts for reward model evaluation, yielding lower average scores but higher correlation with downstream best-of-N sampling and RLHF training performance.
-
WildGuard: Open One-Stop Moderation Tools for Safety Risks, Jailbreaks, and Refusals of LLMs
WildGuard is a new open moderation model and dataset for LLM safety that identifies harmful prompts, risky responses, and refusal rates, achieving SOTA open-source performance and sometimes exceeding GPT-4 while cutting jailbreak success from 79.8% to 2.4%.
-
Long Context Transfer from Language to Vision
Extending language model context length enables LMMs to process over 200K visual tokens from long videos without video training, achieving SOTA on Video-MME via dense frame sampling.
-
From Explicit CoT to Implicit CoT: Learning to Internalize CoT Step by Step
Gradual fine-tuning that removes explicit CoT steps lets GPT-2 Small reach 99% accuracy on 9x9 multiplication and Mistral 7B exceed 50% on GSM8K with no intermediate outputs.
-
Smaug: Fixing Failure Modes of Preference Optimisation with DPO-Positive
DPOP is a new loss function that prevents DPO from lowering preferred response likelihoods and outperforms standard DPO on diverse datasets, MT-Bench, and enables Smaug-72B to exceed 80% on the Open LLM Leaderboard.
-
FragileFlow: Spectral Control of Correct-but-Fragile Predictions for Foundation Model Robustness
FragileFlow formalizes margin-aware error flow and applies spectral control through a calibrated margin buffer and class-wise risk matrix, supported by a PAC-Bayes bound, to enhance worst-class robustness in foundation model adaptation while preserving clean accuracy.
-
STELLA: A Multimodal LLM for Protein Functional Annotation via Unified Sequence-Structure Encoding
STELLA aligns ESM3 bimodal sequence-structure encodings with Llama-3.1-8B text modeling to claim state-of-the-art results on protein functional description prediction and enzyme-catalyzed reaction prediction.
-
Reinforcement Learning for LLM Post-Training: A Survey
A survey deriving a unified policy gradient framework for LLM post-training methods and providing technical comparisons of PPO, GRPO, DPO variants.
- Lessons from the Trenches on Reproducible Evaluation of Language Models