Claims Browser

TRACE-CRC avoids the trajectory undercoverage problems of compact stepwise and adaptive conformal baselines

"TRACE-CRC achieves reliable trajectory-level coverage with substantially smaller uncertainty balls than conservative multi-step corrections, while avoiding the trajectory undercoverage of compact stepwise and adaptive conformal baselines."

Cross-cultural meme transcreation has three core challenges: culture-specific knowledge understanding, intent and tone preservation, and multimodal consistency

"we first provide an explicit task analysis of cross-cultural meme transcreation and identify three core challenges: culture-specific knowledge understanding, intent and tone preservation, and multimodal consistency"

A multi-agent framework with specialized agents can effectively address cross-cultural meme transcreation challenges through cultural adaptation, target text rewriting, revision, and conditional visual adjustment

"we propose a multi-agent framework with specialized agents that are coordinated to address these challenges through cultural adaptation, target text rewriting, revision, and conditional visual adjustment"

The proposed meme transcreation method achieves 33.1% average improvement over the strongest baseline in human evaluation and 60% Top-1 ranking rate versus 26% for the second-best baseline in LLM-as-a-Judge evaluation

"In human evaluation, it achieves the best performance on all four dimensions and delivers a 33.1% average improvement over the strongest baseline, while in LLM-as-a-Judge, it attains the highest Top-1 ranking rate (60% versus 26% for the second-best baseline)"

The remaining bottlenecks in cross-cultural meme transcreation lie in humor reconstruction and image-text alignment rather than simple cultural knowledge gaps

"Our error analysis suggests that the remaining bottlenecks lie in humor reconstruction and image-text alignment rather than simple cultural knowledge gaps, pointing to future work on humor transfer"

Attention magnitude is unrelated to a token's causal contribution to the answer in KV cache eviction

"Using a controlled leave-one-out probe, we find that attention magnitude is unrelated to a token's causal contribution to the answer (Spearman $ρ=-0.004$), challenging the premise behind dominant eviction methods."

TwinKV improves performance for a majority of configurations on two KV cache eviction policies when tested on Qwen3-4B

"On Qwen3-4B, TwinKV improves a majority of configurations for two policies, is near-even for a third, and helps only a minority for a fourth adaptive baseline already near a performance ceiling"

Long-context inference is bottlenecked by KV cache memory footprint, especially for small models

"Long-context inference is bottlenecked by the memory footprint of the key-value (KV) cache, especially for small models under tight resource budgets."

Multimodal foundation models show substantial inconsistencies when handling semantically equivalent queries across different modalities (text vs. speech) and languages (English vs. Arabic)

"it remains unclear whether semantically equivalent queries yield consistent judgments across modality (text vs. speech) and language (English vs. Arabic)"

Modality and language shifts introduce substantial triplet-level inconsistencies in multimodal models that are not fully captured by aggregate accuracy metrics

"modality and language shifts introduce substantial triplet-level inconsistencies that are not fully captured by aggregate accuracy"

Speech input amplifies partial failures in multimodal models compared to text input

"speech amplifying partial failures"

Contrastive instability (the conditional rate at which a model fails to resolve all statements within a triplet) can isolate fragmented reasoning from complete failure in multimodal models

"We define contrastive instability as the conditional rate at which a model fails to resolve all statements within a triplet, isolating fragmented reasoning from complete failure"

Current agent safeguards defined over single trajectories have a fundamental composition failure: against attacks fragmented across iterations, every trajectory-scoped monitor has equal true-positive and false-positive rates regardless of expressiveness.

"against an attack whose evidence is fragmented across several iterations, every trajectory-scoped monitor has a true-positive rate equal to its false-positive rate, however expressive it is, because the evidence it would need never appears in the window it sees"

Carrying a geometrically decaying risk score is insufficient for agent safety because the cooling-off period an adversary must wait is a constant that doesn't grow with the horizon.

"the obvious repair of carrying a geometrically decaying risk score is insufficient, because the cooling-off period a patient adversary must wait is a constant that does not grow with the horizon $N$"

LoopHarness with mediated commits and arbiter detection floor can bound expected unauthorized irreversible actions to a constant independent of horizon N, with part of the bound surviving fully colluding verifiers.

"Under mediated commits and an arbiter detection floor $δ_M$, it bounds the expected number of unauthorized irreversible actions by $B+m-1+m/δ_M$, a constant in $N$, of which the $B+m-1$ term is decided by a model-free rule and therefore survives a fully colluding verifier"

Large language model agents are increasingly being deployed as autonomous loops that repeatedly discover work, plan, execute, and verify outcomes across many unattended iterations.

"Large language model agents are increasingly deployed as autonomous loops. Starting from one human goal, such a system repeatedly discovers work, plans, executes tool calls, verifies outcomes and persists state across many unattended iterations."

A monitor retaining cross-iteration state can perfectly separate true-positives from false-positives for attacks fragmented across iterations, whereas trajectory-scoped monitors cannot.

"a monitor retaining cross-iteration state separates the two perfectly"

Large Language Models remain brittle to real-world shifts in node identifiers and task formulation in graph tasks

"Despite their potential in standardized graph tasks, Large Language Models (LLMs) remain brittle to real-world shifts in node identifiers and task formulation."

Extracting topological structures from noisy text is highly fragile for LLMs, which often overfit to surface patterns

"extracting topological structures from noisy text is highly fragile for LLMs, which often overfit to surface patterns"