Evaluation of two latent reasoning models against controls shows observable latent patterns appear without the proposed mechanisms, have graded causal effects on behavior, and concentrate in structured low-rank directions, arguing that patterns are insufficient evidence for reasoning.
hub
Transactions of the Association for Computational Linguistics9, 160–175 (2021)
13 Pith papers cite this work, alongside 41 external citations. Polarity classification is still indexing.
13
Pith papers citing it
41
external citations · Crossref
hub tools
citation-role summary
background 2
method 1
citation-polarity summary
representative citing papers
Finite-answer projections of continuation probabilities stabilize before the answer is parseable, showing 17-31 token mean lead in delayed-verdict tasks with Qwen3-4B-Instruct.
On 2,520 programming tasks, matched Qwen general and coder models reliably raise Bloom cognitive demand but fail to lower it, so execution skill does not imply educational control.
Authors create a benchmark across discrete/continuous and static/dynamical systems and introduce the Causal Abstraction Error (CAE) metric that reliably distinguishes valid from invalid causal abstractions when it includes faithfulness testing.
PROPEL amortizes solver evaluation with a trained activation probe to optimize task generators toward a target solve rate, raising the share of learnable tasks from ~10% to ~20% in coding and SWE experiments.
Transformer represents but does not causally transmit staged algorithmic intermediates for base-digit extraction, diverging from probe predictions.
LLMs encode accurate but brittle internal beliefs about latent game states and convert them poorly into actions, creating systematic gaps that explain strategic failures.
Language models employ a highly localized shared mechanism for filler-gap dependencies but no unified mechanism for NPI licensing, and activation patching generalizes better than supervised alignment search.
Larger LLMs reproduce constructional productivity via entrenchment in coercion cases with nonce words but fail to use statistical preemption to avoid overgeneralizing semantically plausible but unobserved patterns.
Trait-space drift monitoring detects emergent misalignment checkpoints in 7-9B LLMs with 2.2% FNR, 2.9% FPR and 0.99 AUROC, outperforming PCA and SAE baselines.
Overthinking in medical QA is linearly decodable at 71.6% accuracy yet fixed residual-stream steering yields no correction across 29 configurations, while enabling selective abstention with AUROC 0.610.
Inflectional features stay linearly decodable across all layers while lexical identity weakens with depth in modern transformers.
A review paper that organizes conceptual, practical, and socio-technical open problems in mechanistic interpretability.
citing papers explorer
-
Observable Patterns Are Not Explanations: A Causal-Geometric Analysis of Latent Reasoning Models
Evaluation of two latent reasoning models against controls shows observable latent patterns appear without the proposed mechanisms, have graded causal effects on behavior, and concentrate in structured low-rank directions, arguing that patterns are insufficient evidence for reasoning.
-
When Does a Language Model Commit? A Finite-Answer Theory of Pre-Verbalization Commitment
Finite-answer projections of continuation probabilities stabilize before the answer is parseable, showing 17-31 token mean lead in delayed-verdict tasks with Qwen3-4B-Instruct.
-
From Execution to Education: A Bloom-Aligned Framework for Measuring Educational Control in LLMs
On 2,520 programming tasks, matched Qwen general and coder models reliably raise Bloom cognitive demand but fail to lower it, so execution skill does not imply educational control.
-
Validating Causal Abstraction Metrics on Simulated Complex Systems
Authors create a benchmark across discrete/continuous and static/dynamical systems and introduce the Causal Abstraction Error (CAE) metric that reliably distinguishes valid from invalid causal abstractions when it includes faithfulness testing.
-
Breaking the Solver Bottleneck: Training Task Generators at the Learnable Frontier
PROPEL amortizes solver evaluation with a trained activation probe to optimize task generators toward a target solve rate, raising the share of learnable tasks from ~10% to ~20% in coding and SWE experiments.
-
Represented Is Not Computed: A Causal Test of Candidate Algorithmic Intermediates in a Transformer
Transformer represents but does not causally transmit staged algorithmic intermediates for base-digit extraction, diverging from probe predictions.
-
Why Do LLMs Struggle in Strategic Play? Broken Links Between Observations, Beliefs, and Actions
LLMs encode accurate but brittle internal beliefs about latent game states and convert them poorly into actions, creating systematic gaps that explain strategic failures.
-
Fine-Grained Analysis of Shared Syntactic Mechanisms in Language Models
Language models employ a highly localized shared mechanism for filler-gap dependencies but no unified mechanism for NPI licensing, and activation patching generalizes better than supervised alignment search.
-
Linguistic Productivity in Large Language Models: Models Coerce, but do not Preempt
Larger LLMs reproduce constructional productivity via entrenchment in coercion cases with nonce words but fail to use statistical preemption to avoid overgeneralizing semantically plausible but unobserved patterns.
-
Trait-space Monitoring for Emergent Misalignment During Supervised Finetuning
Trait-space drift monitoring detects emergent misalignment checkpoints in 7-9B LLMs with 2.2% FNR, 2.9% FPR and 0.99 AUROC, outperforming PCA and SAE baselines.
-
Decodable but Not Corrected by Fixed Residual-Stream Linear Steering: Evidence from Medical LLM Failure Regimes
Overthinking in medical QA is linearly decodable at 71.6% accuracy yet fixed residual-stream steering yields no correction across 29 configurations, while enabling selective abstention with AUROC 0.610.
-
Model Internal Sleuthing: Finding Lexical Identity and Inflectional Features in Modern Language Models
Inflectional features stay linearly decodable across all layers while lexical identity weakens with depth in modern transformers.
-
Open Problems in Mechanistic Interpretability
A review paper that organizes conceptual, practical, and socio-technical open problems in mechanistic interpretability.