Papers

TL;DRTop System One models (Jev 93.0%, Clef 91.9-93.0%) were highly accurate on rheumatology exam questions but dropped on diagnostic vignettes (Jev 50.0% vs GPT-6 Sol 76.9%), where Jev's high-confidence answers were only 84.6% correct.

TL;DRFive of six tested decision models give women significantly more favourable yes/no answers (+0.05 to +0.10 log-odds), and recommended confidence gates can raise stereotyped answers in ambiguous BBQ gender questions.

TL;DRPDScope decouples prefill and decode phases with phase-specific expert prediction, scheduling, and zero-copy pipelines, achieving 2.5-10.3x prefill and 2.1-6.1x decode throughput gains on memory-bound MoE inference.

TL;DRProfiling LLM inference shows decode compute time and energy scale linearly with output length, while prefill scales linearly and quadratically with input length, with compute time serving as a reliable energy proxy.

TL;DRPrefill-side decode-length predictors are bounded by inherent autoregressive sampling noise: per-prompt lengths follow a normal law with an affine variance law, so system error dominates model error across predictor designs.

TL;DRGDGraph compiles clinical guidelines into executable decision graphs and combines a fast System One model with deterministic traversal, boosting accuracy from 57.3% to 90.8% across 347 cases while beating GPT-4o baselines at lower cost.

TL;DRA linear probe on agent activations, read before each output is written, catches 82% of leaking calls at a 5% benign flag rate in held-out Qwen3-32B episodes, versus 26% for a zero-shot decision-model monitor.

TL;DRLLM eBay, a single-round marketplace paying sellers via a proper scoring rule against hidden-prompt style signals, cuts convergence steps by 8.5% and reduces style-claim error by 96.5% while recovering 96.3% of Oracle welfare.

TL;DRCAViLD is a latent diffusion framework for continuous conditional generation that builds a condition-aware latent space, internalizes guidance via an EMA teacher, and uses covariance-aware progressive distillation to sample in as few as four NFEs.

TL;DRHALO introduces a human-grounded evaluation of proactive agents across 4 need types, finding that over-proactivity—premature timing or excessive intensity—accounts for 61.20% of rejected responses.

TL;DRSurprise-driven continual learners absorb nearly every planted lie; judges, regression tests, model priors, and evidence routing each help partially, but no acceptance signal alone decides whether a claim is true.

TL;DRdLLM-KVQ quantizes the KV cache of diffusion language models by reconstructing feasible reveal trajectories with TRACE and selectively keeping per-layer K-cache entries in high precision, achieving the highest average score among quantized methods in 4-bit and 2-bit settings.

TL;DRIn a within-subjects study with five healthcare participants, a conversational XAI assistant and a traditional dashboard showed complementary strengths, and all three interviewees recommended integrating both modalities into one clinical workflow.

TL;DRA fast black-box query-based attack finds short universal adversarial suffixes that flip Jev's confident decisions, with one suffix flipping 43% of unseen documents and 90-100% of samples across other tasks.

TL;DRUnder controlled retrieval on ten engines, a distilled GEO Citation Ranker 4B (0.542-0.549) is statistically indistinguishable from Jev and two open 27B decision models (0.575-0.578) on 229 GEO-Bench queries.

TL;DRPrimaCons shows Jev violates cross-primitive consistency systematically, with discrepancies predicting decision failure at ROC-AUC 0.91; the CPRC calibrator cuts inconsistency by 58.57% and boosts accuracy by +2.97%.

TL;DRWorld Reward Models like ForeCritic learn directly from executed program pairs to pick the better program (79.4% on 12,438 held-out pairs), cutting agent task time by 4.4--7.5x without hurting scores.

TL;DRSatOps-Decision applies Jev-style discriminative decision models to satellite operations, achieving 85.3% raw accuracy with 90.3 ms median latency, a 5.47x speedup over Qwen3.5-0.8B generation.

TL;DRJevAct, a runtime verification framework using a prefill-only decision model with executable policy checks, improves tool-using LLM agent reliability across eleven agents on τ^2-bench and BFCL, with mixed results on Airline.

TL;DRAcross 1,273 MedQA questions, 13 decision models were cheaper and faster than 15 generative configurations but reached lower peak accuracy (89.08% vs. 96.86%), with high-probability errors and variable probability repeatability.