Claims Browser

The act/don't-act gate failure can be fixed through training: supervised fine-tuning of a 3B model on 540 synthetic cases drives commitment on unpredictable questions to 0.0% and transfers to three unseen domains

"Because the gate is separable, it can be trained. Supervised fine-tuning of a 3B model on 540 synthetic cases, predominantly dice, coins, jars and timers, drives commitment to 0.0% on the original cases and transfers to three unseen domains."

The trained act/don't-act gate is context-fragile and only holds when response format allows reasoning; rigid formats leave models confident and wrong on questions they otherwise answer correctly

"It does not survive everything: the gate holds exactly when the response format leaves room to reason, and rigid formats that remove that room leave the model confident and wrong on questions it otherwise answers correctly."

Deployment of LLMs needs to account for both the trainability of the act/don't-act gate and its context-fragility

"The gate is trainable and context-fragile, and deployment needs both halves of that sentence."

Video foundation models are beginning to change film and video production

"Recently, video foundation models are beginning to change film and video production"

Magpie enables applying generative models to real-time game rendering while preserving gameplay designability and reproducibility

"Magpie provides a system-level implementation path for applying generative models to real-time game rendering. It preserves gameplay designability and reproducibility, and reduces the dependence of early game prototypes on complete visual assets."

Generative models can reduce the dependence of early game prototypes on complete visual assets

"reduces the dependence of early game prototypes on complete visual assets"

Ancient Chinese artifact text recognition remains fundamentally unsolved despite advances in Vision-Language Models and OCR-specialist models

"Extensive experiments on Ancient-Bench covering general Vision-Language Models (VLMs) and OCR-specialist models reveal that ancient Chinese artifact text recognition remains fundamentally unsolved, with persistent challenges in variant characters, specialized symbols, and hallucination."

Existing benchmarks for ancient Chinese text recognition suffer from fragmentation in temporal coverage, medium diversity, and script type completeness

"However, existing benchmarks suffer from ''fragmentation'', manifested in limited temporal coverage, limited medium diversity, and incomplete script types."

Current models face persistent challenges with variant characters, specialized symbols, and hallucination when processing ancient Chinese texts

"persistent challenges in variant characters, specialized symbols, and hallucination"

Foundation models that use a full year of Earth observations can have their embeddings recomputed over much shorter windows (down to a single day) while retaining substantial value for land-use/land-cover mapping

"Keeping the encoder frozen, we recompute its embeddings over varying observation windows, from a full year down to a single day."

For phenology-separated classes like crop types, Tessera embeddings achieve 46% better performance than the best from-scratch models

"Where classes are separated by phenology, as for the crop types of PASTIS-R, they reach a mean Intersection-over-Union of $58.3$, about $46\%$ above the best from-scratch model."

Tessera embeddings remain markedly more label-efficient than from-scratch models even when classes are temporally stable

"On both datasets, Tessera embeddings remain markedly more label-efficient."

Contracting the observation window from one year to one month costs 39% of segmentation accuracy on crop classification but only 5% on stable land cover classes

"Contracting the window from one year to one month costs $39\%$ of the segmentation accuracy on PASTIS-R but only $5\%$ on DynamicEarthNet."

Single-day embeddings from foundation models can still classify land cover at 3.4 times the chance level

"Single-day embeddings still classify land cover in LUCAS at $3.4$ times the chance level."

Temporal coverage in Earth observation foundation models is a tunable cost rather than a fixed prerequisite, enabling near-real-time mapping and faster refresh cycles

"Our study shows that temporal coverage is therefore a tunable cost rather than a fixed prerequisite, opening regimes such as near-real-time mapping and faster land-use/land-cover refresh cycles."

Existing evaluations of spoken dialogue understanding often allow shortcuts based on transcripts or single-modality solutions, obscuring whether models genuinely ground predictions in speech

"existing evaluations often allow shortcuts based on transcripts or single-modality solutions, obscuring whether models genuinely ground predictions in speech"

Strong text-only LLMs exceed 90% accuracy in consistent cases but drop to 33-48% in conflict cases where acoustic and textual signals disagree

"strong text-only LLMs exceed 90% accuracy in consistent cases but drop to 33-48% in conflict cases"

Direct AudioLLMs provide only partial grounding, still selecting the transcript-biased trap in roughly 30-40% of conflict cases

"Direct AudioLLMs provide only partial grounding, still selecting the transcript-biased trap in roughly 30-40% of conflict cases"