4arXiv cs.CL (Computation and Language)·
Metaphor Tracer: Theory-informed analysis of LLM hidden states for discourse structure
A new arXiv preprint introduces Metaphor Tracer, a method for scoring token positions in a single forward pass on two properties — an 'aggregator' (whether a position consolidates the whole text) and a 'differentiator' (whether other tokens are transiently carried into its subspace) — without any training. The approach is validated across three unrelated models using engineered registers and psychoanalyst-annotated clinical transcripts as ground truth, achieving 34/36 cell matches on the latter. The work argues that structural value is relational (a property of a token's place in a specific text) rather than essentialist (a property of its vector alone), and finds that instruction tuning raises discourse-reading fidelity without affecting type-level transfer.
Related guides (1)
Related events (8)
Causal tracing reveals mid-to-late decoder layers drive LLM stance sensitivity to linguistic construction
A new arXiv preprint investigates whether linguistic construction choices (beyond lexical variation) systematically shift LLM stance judgments, using political stance classification as a case study. The authors create six controlled rewrite types that preserve or invert meaning, then apply activation patching across four open-weight models to localize where these shifts originate. Results show that mid-to-late decoder layers, particularly block outputs at the final prompt position, provide the strongest restoration signal for the original stance distribution. The work extends mechanistic interpretability methods to a socially sensitive domain and highlights a largely overlooked axis of LLM prompt sensitivity.
Mechanistic interpretability study reveals geometric structure of LLM-as-judge bias in hidden states
A new arXiv preprint investigates LLM-as-judge scoring bias at the representation level rather than the input-output level, studying seven judge models across seven bias types and nine benchmarks. The authors find that biased inputs are displaced along low-dimensional, type-specific subspaces in activation space, and that steering hidden states along these subspaces causally controls scoring direction. A linear projection onto bias-direction features predicts judge failures on unseen benchmarks, substantially outperforming text-based alternatives. The work provides a mechanistic account that unifies geometric structure, causal control, and operational prediction within a single framework.
Token probability measurements reveal production-perception asymmetry in LLMs
A new arXiv preprint investigates whether LLMs exhibit a functional analog to the psycholinguistic production-perception distinction, using direct token probability measurements rather than metalinguistic prompting. Using Llama-3.1-8B and four other open-weight models, the authors find that production-perception prompt distances consistently exceed production-production distances by a ratio of ~1.8, with near-ceiling correlations in the production-production control confirming the effect is specific to communicative framing. The effect replicates across five models spanning base and instruction-tuned variants, and temporal analysis shows perception prompts exert strongest influence at sequence beginnings. The findings suggest prompt framing alone induces a production-perception distinction in decoder-only architectures.
Mechanistic analysis of how LLMs encode essay quality in internal representations
Researchers systematically probe the hidden representations of eight LLMs across three essay datasets (ASAP++, CSEE, ENEM) to understand how automated essay scoring (AES) works internally. Using linear probing, dimensionality reduction, and neuron-level analysis, they find essay quality is encoded in a linearly accessible form that emerges progressively across layers and partially transfers across prompts. Individual 'essay scoring neurons' are identified whose activations correlate with scores and respond to targeted interventions, with longer essays relying more on deeper layers. The work contributes to mechanistic interpretability of LLM-based scoring systems.
PropMe framework distinguishes memorization capability from propensity in LLMs
A new arXiv preprint introduces PropMe, a framework that separates whether LLMs can be forced to reproduce training data (capability) from whether they do so under ordinary use (propensity). The authors also release SimpleTrace, a lightweight pipeline using infini-gram to attribute model outputs to training corpora. Evaluating two open models on Common Pile and Dynaword, they find a consistent gap: adversarial prefix attacks elicit strong memorization, but propensity scores remain low in non-adversarial settings. The paper argues memorization audits should report both worst-case extractability and ordinary leakage propensity.
MATCHA: Contrastive Semantic Alignment Metric for LLM Evaluation
MATCHA is a new automatic evaluation metric for LLMs that addresses a fundamental flaw in existing metrics: both token-overlap (ROUGE) and embedding-based (BERTScore) metrics routinely assign near-identical scores to semantically contradictory texts. The metric uses a dual-view approach that rewards proximity to a gold reference while penalizing adversarially generated counterfactual contradictions. Evaluated across eight benchmarks spanning QA, summarization, NLI, and semantic similarity tasks, MATCHA outperforms 23 embedding models and achieves 18.38% and 20.82% improvements over ROUGE-L and BERTScore respectively on TruthfulQA. Code and metric are publicly released.
Survey argues LLMs should be designed with explicit conceptual representations, not just probed for them
A new arXiv preprint proposes treating 'concepts' as a first-class design axis for LLMs, contrasting the current norm of recovering concept-like structure post-hoc (via probing or dictionary learning) with the alternative of deliberately engineering it into training objectives, architectures, or inference procedures. The authors map the design space along two dimensions: pipeline stage and whether structure is internally derived or externally grounded. Key findings include that inference-time approaches are underexplored, related ideas have developed in isolation across pipeline stages, and externally grounded methods span the full pipeline under inconsistent terminology.
Calibrated LLM annotation and encoder transfer for measuring human values in social media text
A new arXiv preprint investigates how different LLMs, prompts, and instruction languages operationalize Schwartz's theory of basic human values when annotating non-English social media posts. The authors evaluate annotation quality beyond standard F1 metrics, examining structural alignment, error structure, and confidence-ambiguity relations, finding that iterative prompt calibration reduces misattributions. They also demonstrate that LLM annotations can be transferred to a smaller encoder model via soft-label training, preserving theory-grounded value interpretations and uncertainty information.