Skip to main content
Search

Korea University

12 related articles

Subscribe to updates
  1. arXiv

    Search interfaces stop being hand-fixed: SearchWeaver rewrites tools and instructions per corpus, lifting answer accuracy from 71.6% to 78.4%

    How well a language-model agent searches depends on the interface it is given, yet existing interfaces are set by hand and applied unchanged to every corpus. SearchWeaver jointly revises corpus artifacts, executable tools and usage instructions using the agent's execution trajectories and task rewards, then selects by downstream search performance. With GPT-5.4-nano, the optimized interfaces raise BrowseComp-Plus answer accuracy from 71.6% to 78.4% and average nDCG@10 across four BRIGHT subsets from 45.7 to 55.0, while cutting tool execution time.

  2. aclanthology.org

    Retrosynthesis stops guessing reactants: an 8B model disconnects bonds first and lifts feasible template diversity from 2.02 to 3.43 on rare reactions

    Retrosynthesis prediction shifts from emitting reactants directly to a four-stage reasoning trace that mirrors a chemist's disconnection strategy: product analysis, key-substructure identification, strategic bond disconnection, and synthetic-equivalent mapping. The model is fine-tuned on structured rationales from SyntheticRetro, then trained with reinforcement learning on a round-trip reward that passes predicted reactants through a forward synthesis model and rewards reconstructions of the original product. On the 500-instance ORDerly main test set, RetroReasoner (RL) reaches Exact@1 of 0.526 and Round-trip@1 of 0.826; on the rare atom/token subset it reaches a feasible ratio of 0.557 and template diversity of 2.97, above the same-architecture Prediction-Only (RL) at 0.478 and 1.97.

  3. arXiv

    An excavator stops taking language orders at face value: a one-use lease ties approval to the exact motion, cutting closed-site dumping from 1.23 to 0.33 cubic meters

    An LLM-directed excavator can execute mechanically feasible motion whose destination or authority no longer matches the operator's approval. The work binds authorization to the exact motion: a one-use lease ties the action, generated motion, mission contract and admission tolerance together, atomic dispatch rechecks current authority and measured joint drift, and revocation requests a carry-pitch quintic stop. Across a 672-case interface suite the gateway admitted all 120 benign cases and produced zero unauthorized dispatches in the attack suite; in synchronous simulation, dump-phase revocation cut maximum closed-site deposition from 1.23 to 0.33 cubic meters.

  4. arXiv

    RL-ARC calibrates answer confidence with reasoning confidence, cutting OOD ECE to 0.16 while holding accuracy

    The work proposes RL-ARC, a calibration-aware reinforcement learning framework that, on top of RLVR, elicits both the model's confidence in its reasoning process and its confidence in its answer, using reasoning confidence as an auxiliary signal: reasoning-guided regularization when the prediction is correct and an overconfidence penalty when it is incorrect; across mathematical reasoning and out-of-distribution complex reasoning and factual QA benchmarks, RL-ARC improves calibration metrics over existing calibration-aware training methods while keeping accuracy comparable to RLVR, which optimizes correctness alone.

  5. arXiv

    RucTangle splits coding-agent patches into commit histories that stay runnable at every commit and lifts pass@1 by 16.6% on average in regression repair

    The work introduces RucTangle, an agentic commit-untangling method that groups a patch's changes by development purpose, orders them, applies each commit and checks whether the code still runs, revising order first, then regrouping via the agent, and finally merging adjacent commits while preserving the patch's final code; on 131 agent-generated patches it is the only method that reconstructs every patch with all commits runnable, at a median of eight commits, whereas baselines leave 20.6%–37.4% of histories with at least one unrunnable version. The authors also present TangleEval, which has two coding agents repair 453 regression-introducing patches; adding RucTangle histories plus git bisect results raises pass@1 from 26.9% to 31.3% and from 30.8% to 36.0%.

  6. arXiv

    SoloQ quantizes diffusion language models to 4 bits without calibration, holding accuracy on LLaDA, Dream, Fast-dLLM v2 and Nemotron-Labs-Diffusion while beating calibration-based baselines

    The authors present SoloQ, which uses a K-RPBH rotation to map weights and activations into a predictable near-Gaussian basis plus a scale-only bias correction for calibration-free 4-bit quantization, and quantizes the KV cache at block commit for block-diffusion models; across LLaDA-Base, LLaDA-1.5, Dream-7B, Fast-dLLM v2 and Nemotron-Labs-Diffusion, SoloQ-C and SoloQ-N retain accuracy under W4A4 and W4A4KV4 and exceed calibration-based baselines such as DLLMQuant++, STaR-Quant and FAIR-Calib on average, with the NVFP4 path also cutting peak memory and speeding up end-to-end inference.

  7. arXiv

    Across three 7B-scale models and three datasets, the hallucination signal is dominated by a single mean-shift direction, and simple L2 logistic regression reaches 0.952 AUROC while bounding or outperforming twelve architectural alternatives

    This work characterizes the geometry of hidden-state probe signals for LLM hallucination detection across three 7B-scale models and three datasets in a paired-example paradigm, finding the signal overwhelmingly dominated by a single mean-shift component whose removal collapses detection to chance, with shrinkage linear discriminant analysis closing about 73% of the gap between 1D and full-dimensional classifiers, a simple L2-regularized logistic regression at 0.952 AUROC bounding or outperforming twelve controlled architectural alternatives, and LayerMix multi-layer aggregation matching oracle-layer performance without oracle access.

  8. arXiv

    CPR routes token-by-token between a base model and its SFT expert, beating the SFT expert by 1.4-5.5% in domain performance while cutting the general-capability drop from 3.4-14.5% to at most 0.5%

    The work proposes CPR (Critical-Point Routing), a token-level routing framework between a base model and its SFT expert derivative that trains a lightweight hierarchical router to estimate the per-token expert-call probability based on critical tokens where the base model fails but the expert succeeds, paired with an inference procedure combining momentum smoothing and threshold gating; across diverse model-domain configurations CPR achieves state-of-the-art in all settings, surpassing the SFT expert by 1.4-5.5% in domain performance while recovering its general-capability drop from 3.4-14.5% to at most 0.5%, with minimal overhead from invoking the expert on only one-third of tokens.

  9. arXiv

    Task-vector variance predicts merge collapse, and PRISM repairs five destructive merges without data

    The work proposes using the variance of specialists' task vectors across specialists as a measure of interference, yielding a pre-merge score that predicts collapse, and designs PRISM: an operator that averages task vectors first and then soft-thresholds each layer at a level set by that layer's interference; across twenty-two merge configurations from four model families only destructive merges exceed the score threshold, twelve of fourteen pre-evaluation predictions were correct, and PRISM keeps all five destructive merges above the threshold within evaluation noise of the base model without data or tuning, where plain averaging falls at least 14.4 points below it or collapses entirely.

  10. arXiv

    RenderRank reranks documents from compressed visual tokens rendered as images, reaching 55.96 average NDCG@10 on BEIR with 16.5–35.5% fewer input tokens

    RenderRank renders document text as images and encodes them into compressed visual tokens, aligning relevance scores to a text-based teacher and then refining positive-versus-negative scores within each query, reaching an average NDCG@10 of 55.96 on 11 BEIR datasets with 290.07 average input tokens, 16.5–35.5% fewer than the evaluated text-based rerankers, and an average NDCG@10 of 88.27 on four long-document datasets at roughly half their average input token count.