Skip to main content
Search

Harbin Institute of Technology

20 related articles

Subscribe to updates
  1. arXiv

    Long-sequence distillation stops handing the teacher the whole answer: ATI-OPD rewrites only the earliest divergent segment, and students' average accuracy on their own beats standard on-policy distillation

    On-policy distillation trains a student on its own responses using token-level teacher feedback, but as responses grow longer the student's prefix drifts from the teacher's own reasoning and later feedback weakens, bringing gradient spikes and extremely long responses. ATI-OPD has the teacher insert a short replacement at the earliest high-disagreement segment, hands generation back to the student, and excludes persistently divergent suffixes from the distillation loss. Across five mathematical reasoning benchmarks and two 1.5B teacher-student pairs, average student-only accuracy exceeds standard on-policy distillation.

  2. bioRxiv

    Tissue proteins without antibodies: MorphoProt lets RNA pick the H&E scale, beating SPRINT on held-out tonsil, spleen and liver

    Spatial transcriptomics often measures RNA without matching protein, and RNA abundance does not fully reflect protein abundance. MorphoProt uses each spot's RNA expression as a query over five H&E fields of view, then aggregates physical and transcriptomic neighborhoods through gated pathways before jointly decoding the protein panel. On held-out sections of human tonsil, mouse spleen and human liver it reached mean protein-wise Spearman correlations of 0.652, 0.604 and 0.372, above RNA-only, graph-based, HIPI and SPRINT comparisons in all three tissues.

  3. arXiv

    Self-distillation's teacher also needs teaching: TeachBack's bidirectional preferences lift Qwen3 math averages by up to 4.3 points

    On-policy self-distillation lets one model act as a problem-only student and as a teacher given a reference solution, but the reference pushes the teacher toward short, answer-conditioned derivations that omit exploration and self-correction. TeachBack injects a trailing slice of the reference reasoning as optional working notes and optimizes preferences in both directions: teacher trajectories should win under the problem-only prompt, while student trajectories reshape future privileged teacher rollouts. Across Qwen3-1.7B, 4B and 8B, math averages rise by 4.3, 2.2 and 2.5 points over base, and Knight-and-Knaves averages by 6.1, 3.9 and 7.5 points.

  4. Typhoon intensity forecasts stop reading cloud images alone: TCFNet steers visual encoding with pressure trends, cutting 24-hour pressure error to 9.52 hPa

    Tropical cyclone intensity forecasting has relied on best-track data or geostationary satellite imagery, yet similar cloud patterns can accompany intensification, steady state or weakening, and generic fusion strategies cannot tell them apart. TCFNet defines intensity trend from pressure gradients, uses it to select positive pairs for contrastive learning so the visual encoder learns cloud representations aligned with intensity change, then propagates states continuously to arbitrary lead times with neural ordinary differential equations. For 24-hour forecasts in the Western North Pacific and South Pacific, pressure RMSE is 9.52 hPa and 7.00 hPa respectively, and the critical success index for rapid intensification is 45.26%.

  5. Breast cancer age-linked genes fall from 80 to four: only IL7R, SQSTM1, EGR1 and IGF1 survive replication and receptor robustness

    Age-associated gene expression signals in breast tumors narrow from 80 to four candidates, IL7R, SQSTM1, EGR1 and IGF1, after receptor and stage adjustment, cross-platform replication and estrogen-receptor stratification. The same molecular workflow is also encoded as a bounded symbolic controller that completes, blocks or fails safe against a prespecified reference policy across 760 scenarios, rather than only finishing. The output is research prioritization, not a clinical biomarker or treatment recommendation.

  6. arXiv

    Mixed-variable experiments stop leaning on structure alone: LLM ranking priors plus structural smoothing reach near-optimal candidates faster on six chemistry and materials tasks

    TagBO treats candidate rankings produced by a large language model as coarse task-aware priors and fuses them with local structural smoothing into a similarity graph, reaching high-quality candidates in fewer evaluations under a fixed budget on six mixed-variable chemistry and materials optimization tasks.

  7. arXiv

    Navigation agents stop re-running finished instructions: coupling progress estimation to decisions lifts R2R success by 3 points

    Vision-language navigation agents previously matched a whole instruction against candidate trajectories, so already executed content kept interfering with later decisions. VLN-AIEP couples instruction-execution progress estimation with navigation decision-making: it uses the visited trajectory to suppress embeddings of executed content, then matches candidate locations only against pending content, so progress can be learned directly from navigation decision labels without explicit progress annotations. Across R2R, RxR, CVDN and R2R-CE, the method reaches competitive results on multiple metrics, including 77% success on R2R val-unseen, 3 points above PRET.

  8. arXiv

    Neural network tuning stops rewriting the whole script: three LLM roles edit only chosen blocks, lifting MNIST accuracy from 80% to 93%

    A multi-agent large language model framework confines neural network tuning to designated functional blocks while the overall network structure stays fixed. Such key components were previously chosen by hand, and existing agent systems either optimize the entire machine learning script or build a network from a text description. With analyst, coder, and recorder roles plus a sandboxed environment and an experiment log, the framework beats its baselines on five tasks: MNIST classification accuracy rises 13 percentage points, digits 6 points, the relative error on the single-soliton generalized nonlinear Schrödinger equation drops about 2491-fold, the conservation error drops about 325-fold in the selected run, and mass drift on the Kuramoto–Sivashinsky problem drops 14-fold.

  9. arXiv

    MoonGS reconstructs lunar terrain from image pairs with a frozen vision foundation model and semantic ranking loss, reaching 23.80 dB PSNR on LuSNAR

    MoonGS is a feed-forward 3D Gaussian Splatting framework for lunar scenes: given only two images it predicts pixel-aligned Gaussian primitives in a single forward pass with no per-scene optimization, using robust depth features from a frozen vision foundation model (DUSt3R, or VGGT), semantic priors plus a semantic ranking loss to regularize weakly textured background depth, and an entropy-guided heuristic resampling strategy to augment sparse observations. On the LuSNAR benchmark and the authors' weak-texture MoonBlender dataset it outperforms feed-forward NeRF/3DGS baselines on PSNR, SSIM and LPIPS while keeping sub-second inference.

  10. arXiv

    D-OPCD distills agent-harness prompt experience into diffusion weights, lifting direct-generation average from 60.52 to 65.09

    The authors propose Diffusion On-Policy Context Distillation (D-OPCD), which treats the agent-harness-produced prompt as privileged teacher context and supervises a student conditioned only on the original query at states along the student's own denoising trajectory, thereby writing prompt-mediated harness gains into the diffusion model's weights; paired with the Auto Skill Evolver (ASE), direct generation improves from 60.52 to 65.09 averaged over four benchmarks, and the updated generator under a skill-free harness slightly exceeds the original skill-equipped agent on average.