arXiv AI
Aug 3
On the Generalization of Steering Vectors for Chain-of-Thought Faithfulness
arXiv:2607. 29062v1 Announce Type: new Abstract: Model capabilities have improved in large part due to scaling chain of thought.
From Geometry to Behavior: How Instruction Exposure Unlocks Latent Rhetorical Directions in LLMs
The paper investigates whether language models possess latent rhetorical abilities, specifically analogy reasoning, before instruction tuning. By extracting an "Analogy Vector" from the residual stream, the authors demonstrate that this vector can steer explanation style across several instruction-tuned models, improving performance metrics. Their findings reveal a continuous gradient of behavioral accessibility tied to the amount of instruction exposure during training, challenging the binary base‑vs‑instruct distinction and suggesting that instruction exposure shapes rhetorical knowledge in activation space.
Related stories
arXiv AI
Sep 2
Do as I Say, Not as I Do: Instruction-Induction Conflict in LLMs
The paper investigates how large language models balance instruction-following with pattern completion when the two objectives conflict. By creating dialogues where a user instruction to act in a target way T is opposed by assistant turns that demonstrate a competing pattern P, the authors measure instruction-following rates across 13 models and 16 instructions over up to 50 turns. Results show wide variability (1%–99%) in instruction adherence, with robustness influenced by instruction content, output format, and chain-of-thought reasoning, but overall instruction-following remains brittle under induction pressure.
arXiv Machine Learning
Sep 22
From Concept Alignment to Causal Grounding: An Intervention Test of Chain-of-Thought Faithfulness
arXiv:2609.23065v1 Announce Type: cross Abstract: Chain-of-thought (CoT) can sound plausible yet be unfaithful to the model's underlying reasoning. Most prior work probes CoT faithfulness through inp...
arXiv Machine Learning
Sep 22
Characterizing Model-Native Skills
arXiv:2604.17614v2 Announce Type: replace-cross Abstract: Skills are a natural unit for describing what a language model can do and how its behavior can be changed. However, existing characterization...
arXiv AI
Aug 3
Metaphor-Induced Algorithmic Steering: Cross-Domain Procedural Transfer in LLM Code Generation
arXiv:2607. 28683v1 Announce Type: cross Abstract: Large language models benefit from elements in natural language, such as metaphors and analogies in training data and inference input to achieve generalisability across different domains.
arXiv AI
Jul 13
GrAInS: Gradient-based Attribution for Inference-Time Steering of LLMs and VLMs
arXiv:2507. 18043v2 Announce Type: replace-cross Abstract: Inference-time steering methods offer a lightweight alternative to fine-tuning large language models (LLMs) and vision-language models (VLMs) by modifying internal activations at test time without updating model weights.
arXiv AI
1d ago
Farther the Shift, Sparser the Representation: Analyzing OOD Mechanisms in LLMs
The paper examines how Large Language Models’ internal representations become sparser as task difficulty rises, finding that harder inputs lead to more concentrated activation patterns in the last hidden state. This sparsity–difficulty relationship holds across various models and domains, indicating that models learn dense representations for familiar patterns but default to sparse activations for unfamiliar ones. Leveraging this insight, the authors propose Sparsity‑Guided Curriculum In‑Context Learning (SG‑ICL), which selects few‑shot demonstrations based on sparsity and outperforms standard CoT and Auto‑CoT baselines on the MATH‑500 benchmark.
arXiv AI
Sep 10
Aha-Flow Distillation: Flow Markers Matter in LLM Reasoning
The paper introduces the Flow Moment, a reasoning pattern marked by sustained, process‑confirming verbalizations, contrasting with the revision‑oriented Aha Moment. It proposes Flow‑CoT, a rewritten version of reasoning traces that preserves content while highlighting Flow Markers, and uses it as auxiliary supervision in on‑policy self‑distillation (OPSD). The authors further present Aha‑Flow Distillation (AFD), a dual‑mode extension of OPSD that pairs concise solution‑based supervision (Aha branch) with rewritten Flow‑CoT under a confident reasoning instruction (Flow branch). Experiments on AIME25 and HMMT25 with Qwen3‑8B and Qwen3‑4B models show consistent performance gains, and controlled ablations confirm that the dual‑mode training structure contributes to the improvement.
arXiv AI
Aug 18
When Do Explanations Help In-Context Learning? A Comparative Study of Natural Language Explanation Types and Faithfulness
arXiv:2608. 16627v1 Announce Type: cross Abstract: Natural language explanations (NLEs) are increasingly used as inputs, for example, as few-shot rationales that influence model behavior in in-context learning (ICL).
arXiv Machine Learning
Sep 23
Rewired or Gated? How Instruction Tuning Shapes Knowledge-Conflict Circuits in LLMs
arXiv:2609.25602v1 Announce Type: new Abstract: In language models, the choice between believing the prompt and believing the weights is made by a handful of identifiable attention heads. Instruction...
arXiv Computation and Language
Sep 1
Detecting Hidden Chain-of-Thought in Large Language Models with Linguistic, Behavioral, and Mechanistic Indicators
arXiv:2608.29956v1 Announce Type: new Abstract: Large language models often answer complex reasoning questions without revealing intermediate steps, raising whether they reason latently or complete p...
arXiv Computation and Language
2d ago
Instruction-Tuned Models Locally Produce More Syntactic Overlap -- But Not More Conditional Reuse -- Than Humans Do
The study examines whether large language models (LLMs) show syntactic convergence with human users by analyzing turn‑adjacent reuse of context‑free grammar (CFG) rules in 1,901 matched positions across sixteen open‑weight Llama and Gemma models. All models exhibited greater CFG‑rule overlap with the preceding human turn than with unrelated primes, especially for lower‑frequency rules, and instruction‑tuned models showed even higher overlap with the actual prime than the human responses they replaced. However, instruction tuning also increased overlap with unrelated primes, reduced the actual‑versus‑random difference, and lowered conditional rule‑reuse odds when controlling for rule‑set size, while consistently producing responses with higher lexical and semantic similarity to the preceding turn.