Ya-Fen Yeh

Which Interference Does 4-Bit Weight Quantization Harm? It Depends on the Operating Point

Language models are commonly served with 4-bit weights, and the format is chosen on perplexity and benchmark averages, which do not vary how strongly the context competes with the answer. A quantized model that fails to resolve such interference can answer fluently with the compe...

Anytime-Valid Stopping for Self-Consistency on Open Answer Sets

Self-consistency samples reasoning chains and returns the most frequent answer, and adaptive variants stop sampling early to save compute. Deployments act on the stopped answer, yet these rules check a posterior or window criterion as draws arrive, so the chance of stopping on an...

Anytime-Valid Stopping for Self-Consistency on Open Answer Sets

Self-consistency samples reasoning chains and returns the most frequent answer, and adaptive variants stop sampling early to save compute. Deployments act on the stopped answer, yet these rules check a posterior or window criterion as draws arrive, so the chance of stopping on an...

Covariate Copies Count as Evidence in Time-Series Foundation Models

Covariate-aware time-series foundation models forecast a target from the related series supplied with it, and which series they receive depends on how a pipeline was built: the same sensor exported twice, a temperature in two units, one weather station joined in for several zones...

Accuracy-Matched Is Not Cost-Matched: Pricing and Repairing Token Inflation in INT4 Reasoning Models

INT4 weight quantization is accepted for a reasoning model when it preserves accuracy, but a reasoning model chooses how many tokens it generates, so an INT4 model that matches BF16 accuracy can still cost more per solved problem. Here we price a compressed reasoner by its genera...

One Percent of the Bank: Adaptive Item Selection Reads Targeted Contamination Up While the Whole Bank Barely Moves

Adaptive testing is entering language-model evaluation as an efficiency device, and its low item exposure is offered as incidental contamination protection (Li et al., 2026b). That argument assumes untargeted leakage, yet a maximum-information test asks predictable items that any...

Removing the Competitor Is Not Enough: Reallocating Attention to the Current Value Restores Recall under Proactive Interference

When one key is updated again and again within a single context, language models answer with a value that has already been overwritten, long before the context window fills. Suppressing attention to the stale value reduces those answers, but no study has held the deletion fixed a...

Reduced-Precision Deployment Can Keep the First Mask and Still Change How Interactive Segmenters Correct

Edge deployments of promptable segmenters are checked by the quality of the first mask or by simulated click counts, yet an interactive tool is used through its corrections. We measure deployment on the correction itself: from the same image, click history and starting mask, we c...

Information Effects Are Not Incoherence: Fixed-Evidence Coherence Tests for Tabular Foundation Models

Tabular foundation models answer every query about a table by in-context inference, so whether their answers to related queries fit one joint distribution is a direct test of whether they can be used as probabilistic models of the table. A recent test of marginalization consisten...

Which Interference Does 4-Bit Weight Quantization Harm? It Depends on the Operating Point

Language models are commonly served with 4-bit weights, and the format is chosen on perplexity and benchmark averages, which do not vary how strongly the context competes with the answer. A quantized model that fails to resolve such interference can answer fluently with the compe...

Removing the Competitor Is Not Enough: Reallocating Attention to the Current Value Restores Recall under Proactive Interference

When one key is updated again and again within a single context, language models answer with a value that has already been overwritten, long before the context window fills. Suppressing attention to the stale value reduces those answers, but no study has held the deletion fixed a...

Selecting GPTQ Damping from the Calibration Set by Nested-Ridge Leave-One-Out

GPTQ-style quantization damps the calibration Hessian by a fixed fraction of its mean diagonal (one percent in the GPTQ authors' code, five percent in a maintained implementation), a choice that ignores how many calibration rows there are. Small tabular, time-series and text-embe...

Covariate Copies Count as Evidence in Time-Series Foundation Models

Covariate-aware time-series foundation models forecast a target from the related series supplied with it, and which series they receive depends on how a pipeline was built: the same sensor exported twice, a temperature in two units, one weather station joined in for several zones...

Covariate Copies Count as Evidence in Time-Series Foundation Models

Covariate-aware time-series foundation models forecast a target from a group of related series, and the covariates supplied to them can be redundant: the same sensor exported twice, a temperature in two units, several nearby weather stations. How these models count such redundanc...

Information Effects Are Not Incoherence: Fixed-Evidence Coherence Tests for Tabular Foundation Models

Tabular foundation models answer every query about a table by in-context inference, so whether their answers to related queries fit one joint distribution is a direct test of whether they can be used as probabilistic models of the table. A recent test of marginalization consisten...

Where 4-Bit Weight Quantization Breaks In-Context Interference Resolution: Layout, Operating Point and Format on a Paired-Associate Battery

Controlled evidence on 4-bit weight quantization and in-context interference is proactive-only, limited to one 4-bit format and behavioural. In 1.5–8B open-weight models we cross proactive and retroactive layouts, with a recency-counterbalanced control, with interference dose, li...

Auditing the Transferability of a Published Quantization Comparison: The Verdict Is a Function of the Slice

Practitioners choose between released 4-bit conversions of a model from published comparison sentences such as "AWQ consistently outperforms ... GPTQ ... (7B-70B)", although the files they serve, the serving stack and the way quality is measured all differ from what the sentence...

Selecting GPTQ Damping from the Calibration Set by Nested-Ridge Leave-One-Out

GPTQ-style post-training quantization adds λI to the calibration Hessian before inverting it, with λ a fixed fraction of the mean diagonal (one percent in the GPTQ authors' code, five percent in a maintained implementation) that ignores how many calibration rows there are. Small...

Inflation-Aware Quantization: Restoring INT4 Reasoning Token-Cost Parity with Distillation

This manuscript is a non-peer-reviewed preprint.

Damping Is a Ridge Penalty: Calibration-Only Selection for Post-Training Quantization under Scarce Calibration

This manuscript is a non-peer-reviewed preprint.