论文库

How multimodal language models should encode visual input remains up for debate. Most models use a pretrained vision encoder, while some recent recipes train the encoder from scratch with the language model or remove it entirely. Vision encoders bundle together three design choices: a visual prior from upstream pretraining, dedicated vision capacity, and pre-fusion placement of that capacity (i.e., before the multimodal backbone). Because published models also differ in data, architecture, and compute, comparing them cannot isolate the effect of each choice. We disentangle these choices via a controlled framework where we pretrain all models from scratch on common data and compare them on visual question answering and language benchmarks at matched lifetime compute. We find that the visual prior is not necessary; an encoder trained jointly from scratch is on par with a pretrained one, without a separate training stage. What matters instead is dedicated vision capacity: an encoder-free model trails on visual understanding even when its backbone is enlarged by the encoder's parameter count. Yet that capacity need not precede multimodal fusion, as vision-only modules added in-fusion to early backbone layers perform comparably to a pre-fusion encoder and can lower time to first token. These results also hold across backbone sizes, when encoder pretraining is treated as free, and when starting from a pretrained language model. Overall, our findings favor dedicated vision capacity trained jointly with the backbone and placed early within it.

Coding agents tackle successive tasks on an evolving repository, carrying forward edits and writing memory before the next task is known. Selecting memory for unknown next tasks requires balancing future usefulness against the costs of writing and reading it. The actor must also decide when acting on available memory saves effort as the task unfolds. We introduce Coding-Repository Memory (CoReMem), a cost-internalizing framework that guides memory design and actor adaptation through measured memory value on the next task. Prospective segment valuation compares runs of the next task with and without individual trajectory segments from the same inherited repository state, guiding a compact memory built around anchors: repository locations and executable checks. Actor adaptation estimates on-policy memory value within rollout groups with and without memory, using a reward that prioritizes resolution and penalizes step count in successful runs. Provenance-gated value credit assigns the additional value term to steps identified as using memory-supplied anchors, while every step retains outcome credit. On ChainSWE, CoReMem raises the 27B actor's resolve rate from 21.0% to 25.7%, with 17% fewer steps and 11% fewer total cache-aware tokens, including memory-module calls, relative to the untrained actor model without memory.

Large language models (LLMs) show growing promise in specialty medicine, yet existing evaluations largely focus on isolated clinical encounters rather than assessing the distinct clinical responsibilities along longitudinal care pathways. We introduce NPC-RubricBench, a phase-aware benchmark for evaluating multistage clinical reasoning across the cancer care continuum, using nasopharyngeal carcinoma as a defined disease context. Curated from real-world records of 405 patients at two cancer centers, it comprises 941 evaluation units across six sequential care stages, from work-up through follow-up, and 17,688 patient-specific criteria across five dimensions. Each assessment is restricted to clinical evidence available at the corresponding decision point. Across 30 general-purpose and medical LLMs, we find strong diagnostic scores did not ensure reliable downstream care. Decision Quality ranged from 56.8 to 94.7 in diagnosis on a 0-100 scale, but dropped to 8.9-56.0 in staging and spanned 32.4-94.2 in treatment. Crucially, failure modes differed across phases: explicit decision errors predominated in anatomical staging (48.2%), whereas decision omissions were more frequent during initial work-up (33.4%) and post-treatment follow-up (30.4%). Varying query granularity revealed a trade-off between decision quality and other clinical requirements. Broad open-ended requests reduced decision quality across all 30 models, whereas more detailed, checklist-like prompts improved decision quality but were accompanied by lower evidence faithfulness and safety adherence. These findings highlight the need to assess clinical reliability against the distinct clinical requirements of each phase along the longitudinal care pathway. Finally, NPC-RubricBench informs the design of stage-aware clinical guardrails and offers a reusable blueprint for evaluating LLMs across other cancers and chronic diseases.

Hallucinations remain a major obstacle to deploying large language models (LLMs) in knowledge-intensive settings, where generated responses must be faithfully grounded in provided evidence. Reinforcement learning (RL) is a promising direction for hallucination mitigation, but response-level faithfulness rewards suffer from a granularity mismatch: localized hallucinations can cause supported content to receive spurious penalties. Although recent work introduces fine-grained feedback such as claim-level verification and token-level rewards, unbalanced credit assignment can still induce length, verbosity, or optimization-noise biases. We propose BALTO, a Balanced Token-level Policy Optimization framework for hallucination mitigation. BALTO extracts checkable factual claims, verifies them against the reference context, and projects claim-level judgments to token-level labels. A balanced token-level credit assignment mechanism is introduced into the framework. This design redistributes probability mass from unsupported content toward faithful content, rather than suppressing the entire response. We systematically analyze the limitations of response-level rewards from a theoretical standpoint, and prove BALTO's advantages in training stability and optimization efficiency for hallucination mitigation. Experiments on ConFiQA, RAGTruth, and FinLLM-Eval show that BALTO achieves the highest faithfulness across all six model--benchmark settings and consistently outperforms existing post-training baselines in Q-Score, demonstrating a stronger faithfulness--informativeness trade-off. The code is released at https://github.com/NingLi670/BALTO.

Neural theorem proving increasingly uses retrieved premises and project context to support proof generation. Evaluating these resources requires measuring the advantage of target-selected declarations over comparable alternatives and how this advantage depends on training. We introduce LeanContext, a benchmark of 129,826 Lean whole-proof generation problems, to investigate these questions. We group selected complete preceding mathematical declarations as Mathematical Context (MC) and the remaining selected local source commands as Environment Context (EC). Four aligned input views support crossed training and inference evaluations with shared targets and original Lean verification environments. For both diffusion- and autoregressive-based models, selected declarations outperform comparable alternatives by a larger margin when training includes MC. Final-input audits identify omitted declarations referenced by the target. Prioritizing these dependencies improves Pass@32 by 1.40% and 1.70% for LLaDA and Llama, respectively, on 2,000 targets. The selection experiment operates without access to target reference proofs and holds checkpoints, EC, and inference budgets constant. The code and dataset are available at https://github.com/haaaardcvZ/LeanContext.

AI agents are playing an increasingly important role in physics research. However, existing physics benchmarks only evaluate fixed questions, isolated sub-tasks, or paper reproduction with given solutions, and rarely check research quality alongside correctness. We introduce FrontierPhysics, an end-to-end benchmark with 56 tasks made by 42 physicists based on their own research, covering nine domains across theoretical, computational, and experimental physics. Each task requires agents to review the literature, develop a research plan, conduct research and experiments, and produce structured outputs and scientific manuscripts. Combining deterministic verifiers with expert-authored rubrics, FrontierPhysics evaluates whether agents' research is correct and publishable. Across 22 models in 26 configurations, no configuration passes more than 46 of its 168 rollouts. Manual review of agent trajectories shows that frontier agents' main weaknesses are insufficiently rigorous reasoning and limited expert intuition, rather than gaps in physics knowledge or execution errors.

AI research agents automate the cycle of proposing, implementing, and evaluating experiments, opening a path toward recursive self-improvement. Yet their ability to propose experiments outpaces their capacity to execute them in real environments, making outcome prediction a key capability for sustained self-improvement under limited experimental budgets. We investigate language models as Research World Models (RWMs), which predict the outcomes of candidate interventions across research environments. Our evaluation draws on over 2,600 experimental records from nine research environments spanning pretraining, post-training, and inference, representing more than 171,000 H100 GPU-hours of experimentation. Research knowledge acquired from real experimental experience improves RWM predictions of unseen interventions within the same environment (Spearman +0.27), and can be reused across environments. For example, using only pretraining experience from OLMo3, Marin, and Nanochat, an RWM reduces selection regret in the Qwen3 environment by 78% compared with zero-experience setting. These benefits extend to multi-round Autoresearch under a fixed selection budget: RWMs with in-env and cross-env research knowledge increase the best gain achieved by 15.8% and 11.6%, respectively. Ablations across 13 language models used as RWMs show that adding research knowledge can improve intervention ranking more than changing models or increasing reasoning effort alone. These findings support language models as RWMs and motivate accumulating experimental data for future RWM training.

Efficient preconditioning plays an important role in large-scale first-order optimization. In this paper, we propose an adaptive accelerated optimistic dual averaging ($\text{A}^2$ODA) method for unconstrained convex optimization that adopts low-cost preconditioning based on $p$-norm geometry for $p\in[2,\infty]$. The method is line-search-free and employs a new stepsize selection technique based on an aggregated affine lower model. Its update structure is simple and unified across $p\in[2,\infty]$, with the case $p>2$ requiring no more elaborate updates than the Euclidean case $p=2$. Under a generalized local smoothness condition, we derive a convergence bound that separates the effects of constant and gap-dependent curvature, with the latter potentially yielding faster-decaying terms. In the special case of local smoothness, this bound recovers a nearly optimal convergence rate.