probing

This paper probes character-level transformers to investigate whether they encode the Spanish L-shaped morphome, an irregular morphological pattern, as an abstract class or just surface alternations. The authors find that the encoding is item-specific and localized, but does not generalize like human learners.

This paper investigates how LLMs encode contextual truth — propositions whose truth is determined by in-context evidence — as linear directions in activation space, showing that these representations persist across output policies and can be shifted by partner assertions, with evidence linking representational shifts to sycophantic behavior.

This paper investigates whether a frozen looped transformer can read its own computation quality (pre-answer prediction reaching AUROC 0.797) and whether external interventions can improve outcomes, finding that no tested frozen intervention produces a validated capability gain, a property termed operational proto-introspection.

This paper introduces a white-box instrument using hidden deterministic finite automata to separately measure reward success and latent-state learning in reinforcement learning agents, finding that high reward does not imply task understanding.

This paper investigates how language models under coding agents internally represent the evolving program during iterative editing. The authors find that linear probes can decode program properties (e.g., parsing, test pass rate) from residual streams, and surprisingly, these representations predict future outcomes before the agent materializes edits, revealing a 'latent programming horizon'.

This paper investigates whether pre-trained code models encode type information in their hidden states, using linear probes on Java and Python examples. Results show cross-lingual type representations emerge even from untyped code, and these representations are robust to lexical perturbations and syntactic variations.

This paper probes internal representations of LLM forecasters to improve calibration and assess faithfulness of chain-of-thought reasoning, finding that probes achieve better calibration and act as lie detectors.

This paper investigates whether activation patterns in Polish Bielik LLMs can detect entity familiarity before generation, using unsupervised dispersion measures that achieve near-perfect separation of known vs. fabricated entities across model scales, while showing that factual reliability improves sharply with scale but is harder to predict from activations.

This paper proposes a recall-controlled abort cascade that uses lightweight probes on LLM agent internal representations to detect and abort doomed episodes early, saving up to 47% inference compute while maintaining high recall of successful episodes.

This paper proposes replacing the proprietary GPT-3.5 metadata extractor in Multi-Meta-RAG with a lightweight, deterministic probe trained on hidden states of a small open-source model. The probe achieves 90.9% accuracy, outperforming GPT-3.5 (80.9%) and a substring baseline (88.0%), while avoiding allow-list drift and API costs.

This paper presents a case study using unsupervised articulatory probing to examine how self-supervised speech models encode phonetic features across Mandarin sub-dialects, finding that salient features like labiality remain stable while finer spectral distinctions show dialect-dependent variation.

MemTrace is a benchmark that evaluates LLM agent memory at the knowledge point level, probing how facts behave under varying memory age, question type, and evidence conditions. It reveals that pooled accuracy hides distinct failure modes, and that the main bottleneck is evidence use rather than retrieval.

This paper investigates why instruction-tuned language models give different answers to causal reasoning questions when variable names are replaced with placeholders, finding that the issue stems from representational misalignment rather than information loss. The authors introduce Vernier, a method using paired-view weight updates and mechanism inspection to reveal that answer-relevant content is still present in the placeholder view but misaligned.

This paper introduces 'fragility', a complementary metric to probe accuracy that measures activation-noise level at which probe accuracy collapses, enabling analysis of representation evolution during LLM pre-training even after accuracy saturates.

The article introduces a technique that extracts hidden states from an LLM at the last prompt token to perform classification without text generation, using a small MLP to read the model's internal decision, enabling fast and cheap zero-shot classifiers.

This paper investigates whether open-source quantized LLMs encode a linearly separable truthfulness signal in their hidden states. Across three 7B-8B instruction-tuned models, a linear probe on a single mid-network layer achieves 0.904-1.000 AUROC on hallucination detection benchmarks, outperforming sampling-based methods.

This paper introduces the readout-mediator angle to demonstrate that linear probes can decode information from language model activations that is orthogonal to the model's actual causal computation, undermining probe-based interpretability. The finding replicates across model scales and families, revealing a fundamental failure mode in using probes for mechanistic understanding or safety monitoring.

This paper investigates whether large language models encode syntactic abstractions like phase boundaries that are not captured by Universal Dependencies, using structural probes on wh-movement stimuli with invariant UD distances, finding evidence across 13 LLMs for phase-structure representations that are causally active.

This paper presents a systematic frozen-feature probing study comparing vision-language models (VLMs) and video generation models (VGMs) on spatial intelligence tasks. It finds that VLMs excel at semantic tagging and instance grouping, while VGMs provide better dense geometry and camera motion signals, and a naive fusion of both yields strong performance across all axes.

This paper proposes a Polar Probe that linearly recovers semantic structures from LLM activations by representing entity relations through distance and direction in a learned subspace. Testing across arithmetic, visual scenes, family trees, metro maps, and social interactions shows the code emerges in middle layers, generalizes to new entities, and causally influences model predictions.