Tag
The paper proposes PAIR (Phase-Aligned Intra-question Reasoning), which treats LLM reasoning paths as phase-structured trajectories and evaluates path quality only within the same question, showing that standard correctness probes partly rely on question-level variation and that phase-wise steering yields causal evidence of path-quality directions.
Research demonstrates that a small set of internal coordinates from LLM hidden states can predict future states and enable targeted edits, with prediction error reduced by 69-76% compared to baseline.
The study investigates whether hallucination detection probes trained on monolingual text transfer to Hinglish, finding robust performance and higher hallucination rates in Hindi and Hinglish compared to English.
This paper proposes Bypass Observation, a non-intrusive architecture for extracting intermediate semantic information from large language models to improve interpretability and safety auditing.
ForeSight is a framework that predicts harmful outputs from large language models by analyzing first-token hidden states, improving early risk monitoring efficiency.
The paper proposes a reference-based hidden-state auditing method to detect bias shifts in LLMs with minimal compute, showing high correlation with output-level bias changes across benchmarks.
The article discusses hidden states that an AI agent should track when diagnosing CI failures, such as flaky tests, real bugs, and configuration errors, and seeks feedback on weaknesses and missing states.
OmniLens is a scalable lens method for interpreting LLM hidden states, using low-rank translators and Subset-KL to reduce parameters and memory, enabling a dense ensemble of 482 lenses on LLaMA-3.3-70B at substantially lower cost.
This paper identifies a 'representation confidence gap' in diffusion language models: internal states detect input noise accurately but reported confidence stays high and answer ranking degrades under noise. It introduces a lightweight, training-free extraction tool that leverages hidden states to improve ranking without modifying the base model.
This paper introduces Prompt Embedding Probes (PEP), a parameter-efficient extension of linear probes that uses learnable prompt embeddings on hidden states to detect hallucinations in frozen LLMs. Evaluations on TriviaQA, GSM8K, and MedQA with Qwen3 models show improvements over standard linear probes, including in pre-generation and cross-model settings.
This paper introduces Right Reset (RR), a prefix-removal probing method that uses causal language model hidden-state preservation to identify chunk boundaries in flattened text, recovering 47.7% of original records compared to 25.9% for a BGE baseline.
This paper introduces ParamBench, a difficulty-graded benchmark for LLM tool-call parameter generation, and proposes probe-guided training methods (PBT and PGR) that improve exact-match accuracy from 19.7% to 59.6%.
This paper investigates how the choice of source activations influences activation steering in language models, finding that execution-boundary states (where the model is about to produce target behavior) yield stronger signals, and introduces tail subtraction to improve steering stability.
Introduces SPARK, a method that uses length-controlled hidden-state susceptibility to diagnose and steer reasoning states in LLMs, improving accuracy on mathematical reasoning benchmarks such as GSM8K and MATH-500.
This paper investigates how multimodal instruction-tuning reorganizes the geometric encoding of identity-specifying prompts in transformer hidden states, finding a shift from direction-based to magnitude-based encoding after instruction tuning.
This academic paper introduces finite-lag operator geometry for analyzing recurrent neural network hidden states, deriving a source-centered transport tensor and antisymmetric coordinate circulation to capture directed flow and deterministic recurrent motion beyond static snapshots.
This paper studies the geometric properties of chain-of-thought trajectories in the hidden state space of transformers, introducing effective dimension and kinematic features to predict task hardness and solution correctness from early tokens.
This post reports an observation that reading a long, structured text before answering alters a model's later responses, with behavioral evidence from Claude and mechanistic analysis on open-weight Gemma models showing separable hidden states and sharper probability distributions in instruction-tuned variants.
An empirical study investigating how long, semantically dense benign text can shift a model's latent space trajectory, diluting initial system prompts and bypassing post-training alignment constraints, as observed in both closed and open-source models.
Proposes ReLAR, a reinforcement-guided latent refinement framework that iteratively updates hidden representations in LLMs before decoding, improving reasoning reliability and efficiency compared to chain-of-thought methods.