Skip to main content
Search

University of Pennsylvania

23 related articles

Subscribe to updates
  1. arXiv

    Cascades in LLM agent networks become predictable, but only if each agent is measured with its follower count

    On social networks of large-language-model agents, whether a post spreads depends on how many of each agent's contacts have shared it. Earlier predictions measured a single response curve independent of contact count, or one response in a large fully connected group. Measuring the response separately at each contact count and feeding it into mean-field cascade theory reproduces the cascade sizes and tipping seed fractions of full LLM simulations, while responses that ignore contact count mispredict cascades increasingly on heterogeneous networks.

  2. Multimodal explanations without model internals: MultiSHAP scores each image-patch and text-token pair

    Multimodal models answer questions and retrieve images, yet which image region and which word jointly push a given output score has been hard to quantify. MultiSHAP applies the Shapley Interaction Index from cooperative game theory to estimate, for every image-patch and text-token pair, a second-order synergy or suppression score, using only a query interface that allows separate perturbation of image and text and returns a task-relevant scalar. On 500 VQAv2 samples, deleting patches in the resulting ranking lowered the target-answer probability by 0.1720 on average, versus 0.0032 for the lowest-ranked patches.

  3. arXiv

    Neural decoding stops needing only paired recordings: BeeMO gains cross-task transfer from 6.5 hours of unpaired behavior

    Neural motor decoding has long depended on invasive paired neural-behavioral recordings, while unpaired behavioral data is cheaper and easier to scale. BeeMO decouples the behavior decoder from the neural encoder: the decoder first learns movement dynamics from behavioral trajectories, and neural activity enters through cross-attention only when available. Across intracortical motor datasets in nonhuman primates, adding 6.5 hours of unpaired behavior made BeeMO-POYO significantly better than POYO on random-target reaching and on isometric wrist, and behavior-only pretrained BeeMO-POYO matched POYO pretrained on 32 neural sessions.

  4. A personal reference curve for brain wiring: BRIDGE cuts brain-age error by 3%–9% and ties deviation scores to ADOS in autistic boys

    Brain-connectivity differences in autism and other neurodevelopmental disorders have mostly been read from group-average case-control contrasts, which flatten individual variation. BRIDGE combines a graph variational autoencoder with normative modeling to learn an age-varying reference trajectory of structural connectomes from 1,164 neurotypical participants, then reports a global deviation score from brain-age difference and an individualized regional deviation map from regional embeddings. In 229 autistic children, the global score correlated with ADOS (Spearman R=−0.291) and IQ (Pearson R=0.195) in males, and brain-age prediction error was 3%–9% lower than comparable deep-learning models.

  5. arXiv

    Finer process rewards are not automatically better: decaying the weight lifts long-horizon trailer editing by 45.4%, while a fixed weight truncates 44.1% of rollouts

    On four trailer-editing task types built from 49 real films and their released trailers, decaying an LLM-judged turn-level process reward to zero within the same GRPO setup gives a 45.4% relative gain over outcome-only GRPO on Fine Cut, while the fixed-weight variant scores below outcome-only GRPO on the three shorter task types and truncates 44.1% of rollouts. Process rewards do encourage visual inspection and reasoning early, but continued optimization drives redundant tool calls without a committed result; token-level on-policy self-distillation proves highly sensitive to the teacher's privileged information.

  6. Retrained on Nigerian scans, a brain-age model reaches 9.82 years of error, then 8.78 on a second cohort

    Brain-age models trained mostly on Global North data lose accuracy when applied to Nigerian adults. A population-specific Gaussian process regression model trained on clinical T1-weighted MRI from 309 Nigerian adults reached a mean absolute error of 9.82 years on a 47-participant held-out test set without age-bias correction; on the 44-participant head-to-head set it reached 10.24 years, against 10.52 years for PyBrainAge, 12.33 for SynthBA, 17.51 for ENIGMA-Brain Age and 7.87 for SPARROW. After correction it reached 6.39 years internally but 8.78 years on a separate Nigerian cohort, where every model's R2 was negative.

  7. arXiv

    One bitstream diffusion recipe spans images, speech and proteins: 3.5% word error in joint speech-text, but two fold-level protein figures disagree

    Continuous bitstream diffusion writes images, speech and proteins as binary codes and, with one training and sampling protocol, reaches 3.5% word error rate in joint speech-text generation, while the fold-level agreement of generated protein sequence-structure pairs is reported as 75% in the results section and 69% in the conclusion.

  8. arXiv

    Same attention heads, new task, new role: one internal activation edit changes which fact the model reads and how it influences later layers

    Across five open-weight models, holding graph facts fixed and changing only the task instruction redistributes how the same attention heads influence later computation; in Gemma-2-9B and Qwen3-14B, editing one early internal activation vector switches which fact the model reads while the written instruction and weights stay fixed.

  9. Lesion segmentation drops the per-lesion box: one concept prompt lets SAM3 outline 11 lesion types across 13 datasets

    SAM3 segments lesions from text or image-exemplar concept prompts across 13 datasets and 11 lesion types, beating SAM2 on Dice in 10 datasets, by 4.41 on ISLES22 and 5.39 on ISIC2018.

  10. arXiv

    Production AI optimization stops reading only failures: CoCoEvolve hits 92.8% hard accuracy on DABStep

    Evolutionary optimization for production AI systems no longer picks one candidate from the search frontier and rewrites it from its failures alone. CoCoEvolve has an agentic mutator retrieve evidence from the entire evolutionary history and probe the live system before proposing changes. On DABStep it reaches 92.8% held-out hard accuracy, against 78.4% for Meta-Harness and 30.4% for GEPA under matched conditions; on data-eng-bench it lifts test-set solve rate from 59.1% to 77.4%, past the 73.6% best baseline.