paperlist.ai

AI 顶会论文

AI 顶会论文库

检索 30+ 个顶会的论文，包括 NeurIPS、ICLR、ICML、CVPR、ACL、AAAI 等。输入关键词筛选，打开任意一篇查看摘要、作者和链接。

arXiv

Mixture of Multicenter Experts in Multimodal AI for Debiased Radiotherapy Target Delineation

Yujin Oh, Sangjoon Park, Xiang Li, Pengfei Jin, Yi Wang, Jonathan Paly, Jason Efstathiou, Annie Chan, Jun Won Kim, Hwa Kyung Byun, Ik Jae Lee, Jaeho Cho, Chan Woo Wee, Peng Shu, Peilong Wang, Caiwen Jiang, Nathan Yu, Jason Holmes, Jong Chul Ye, Quanzheng Li, Wei Liu, Woong Sub Koom, Jin Sung Kim, Kyungsang Kim

Clinical decision-making reflects diverse strategies shaped by regional patient populations and institutional protocols. However, most existing medical artificial intelligence (AI) models are trained on highly prevalent data patterns, which reinforces biases and fails to capture the breadth of clinical expertise. Inspired by the recent advances in Mixture of Experts (MoE), we propose a Mixture of Multicenter Experts (MoME) framework to address AI bias in the medical domain without requiring data sharing across institutions. MoME integrates specialized expertise from diverse clinical strategies to enhance model generalizability and adaptability across medical centers. We validate this framework using a multimodal target volume delineation model for prostate cancer radiotherapy. With few-shot training that combines imaging and clinical notes from each center, the model outperformed baselines, particularly in settings with high inter-center variability or limited data availability. Furthermore, MoME enables model customization to local clinical preferences without cross-institutional data exchange, making it especially suitable for resource-constrained settings while promoting broadly generalizable medical AI.

arXiv

Amortizing intractable inference in diffusion models for vision, language, and control

Siddarth Venkatraman, Moksh Jain, Luca Scimeca, Minsu Kim, Marcin Sendera, Mohsin Hasan, Luke Rowe, Sarthak Mittal, Pablo Lemos, Emmanuel Bengio, Alexandre Adam, Jarrid Rector-Brooks, Yoshua Bengio, Glen Berseth, Esmeralda S. Whitammer

Diffusion models have emerged as effective distribution estimators in vision, language, and reinforcement learning, but their use as priors in downstream tasks poses an intractable posterior inference problem. This paper studies amortized sampling of the posterior over data, $\mathbf{x}\sim p^{\rm post}(\mathbf{x})\propto p(\mathbf{x})r(\mathbf{x})$, in a model that consists of a diffusion generative model prior $p(\mathbf{x})$ and a black-box constraint or likelihood function $r(\mathbf{x})$. We state and prove the asymptotic correctness of a data-free learning objective, relative trajectory balance, for training a diffusion model that samples from this posterior, a problem that existing methods solve only approximately or in restricted cases. Relative trajectory balance arises from the generative flow network perspective on diffusion models, which allows the use of deep reinforcement learning techniques to improve mode coverage. Experiments illustrate the broad potential of unbiased inference of arbitrary posteriors under diffusion priors: in vision (classifier guidance), language (infilling under a discrete diffusion LLM), and multimodal data (text-to-image generation). Beyond generative modeling, we apply relative trajectory balance to the problem of continuous control with a score-based behavior prior, achieving state-of-the-art results on benchmarks in offline reinforcement learning.

arXiv

Transformer-Based Autonomous Driving Models and Deployment-Oriented Compression: A Survey

Juan Zhong, Yuhang Shi, Zukang Xu, Xi Chen

Transformer-based models are becoming a central paradigm in autonomous driving because they can capture long-range spatial dependencies, multi-agent interactions, and multimodal context across perception, prediction, and planning. At the same time, their deployment in real vehicles remains difficult because high-capacity attention-based architectures impose substantial latency, memory, and energy overhead. This survey reviews representative Transformer-based autonomous driving models and organizes them by task role, sensing configuration, and architectural design. More importantly, it examines these models from a deployment-oriented perspective and analyzes how efficiency constraints reshape model design choices in practice. We further review compression and acceleration strategies relevant to Transformer-based driving systems, including quantization, pruning, knowledge distillation, low-rank approximation, and efficient attention, and discuss their benefits, limitations, and task-dependent applicability. Rather than treating compression as an isolated post-processing step, we highlight it as a system-level design consideration that directly affects deployability, robustness, and safety. Finally, we identify open challenges and future research directions toward standardized, safety-aware, and hardware-conscious evaluation of efficient autonomous driving systems.

arXiv

Diffusion models as plug-and-play priors

Alexandros Graikos, Esmeralda S. Whitammer, Nebojsa Jojic, Dimitris Samaras

We consider the problem of inferring high-dimensional data $\mathbf{x}$ in a model that consists of a prior $p(\mathbf{x})$ and an auxiliary differentiable constraint $c(\mathbf{x},\mathbf{y})$ on $x$ given some additional information $\mathbf{y}$. In this paper, the prior is an independently trained denoising diffusion generative model. The auxiliary constraint is expected to have a differentiable form, but can come from diverse sources. The possibility of such inference turns diffusion models into plug-and-play modules, thereby allowing a range of potential applications in adapting models to new domains and tasks, such as conditional generation or image segmentation. The structure of diffusion models allows us to perform approximate inference by iterating differentiation through the fixed denoising network enriched with different amounts of noise at each step. Considering many noised versions of $\mathbf{x}$ in evaluation of its fitness is a novel search mechanism that may lead to new algorithms for solving combinatorial optimization problems.

arXiv

Automated Researchers Can Mitigate Well-characterized Alignment Failures

Chen Yueh-Han, Jiaxin Wen, Jan Hendrik Kirchner

Automating alignment research may accelerate progress toward aligned AI, but whether it does is hard to measure. Luckily, many alignment failures, such as deception, sycophancy, and jailbreaks, are already measurable by public benchmarks. We study whether automated alignment researchers (AARs) can post-train to mitigate alignment failures by proposing training methods and data to simultaneously optimize multiple safety benchmarks, while largely preserving general capability. Across 10 alignment failures, the strongest AAR methods significantly reduce the targeted alignment failures and generalize to a held-out benchmark, multi-turn behavioral audits, and models up to 4.7x larger than the target model. As a human baseline, 28 experienced researchers receive up to eight hours to develop one-shot methods for the same benchmarks, but their methods underperform the best AAR methods. Using human ideas as the AARs' initial research direction does not improve performance, suggesting current AARs may not need guidance from experienced researchers. These results suggest that automating alignment research on well-characterized failures may be practical in the near term.

arXiv

VocalAffectBench: Evaluating Vocal Emotion Recognition in AI Audio Models

Luc Debaupte, Models Luc Debaupte, Tyler Baumgartner, Brandon Tai, Candice Fan, Bill Wang, Yi Zhong

Voice products increasingly need affective cues that are present in speech but absent from transcripts. We introduce VocalAffectBench, a public, test-only benchmark for evaluating whether AI audio models can identify expressed vocal emotion from raw audio. The benchmark contains 273 human-recorded English WAV clips from 51 speaker accounts totaling 1.95 hours across seven labels: angry, disgusted, fearful, happy, neutral, sad, and surprised, with 39 clips per class. All baselines are evaluated from audio alone, without transcripts or contextual metadata. Across six released baselines, average accuracy is 35.5%. The strongest baseline, gemini_3_5_flash, reaches 46.5% on the seven-way task, above the 14.3% random baseline but far from robust emotion recognition. A secondary valence-bucket analysis maps labels into positive, neutral, and negative classes, excluding surprised because its valence is ambiguous. Aggregate accuracy under this coarser view is 50.9%. Performance is highly uneven across classes. By recall, neutral is identified most reliably at 75.6% averaged across baselines, while surprised and fearful reach only 10.7% and 15.4%, respectively. These results show that the evaluated baselines can extract some affective signal from speech, but discrete expressed-emotion recognition remains fragile, especially for non-neutral emotions that are often most important in voice agent workflows.

arXiv

The Hallucination Signal Is a Mean Shift: Why Simple Probes Suffice

Jungseob Lee, Jaehyung Seo, Heuiseok Lim

Hidden-state probes effectively detect LLM hallucinations, but the geometry of the signal remains poorly characterized, driving increasingly complex probe architectures. Across three 7B-scale models and three datasets in a paired-example paradigm, we find the signal overwhelmingly dominated by a single mean-shift component, and removing this direction collapses detection to chance. Shrinkage linear discriminant analysis closes about 73% of the gap between 1D and full-dimensional classifiers, so apparent architectural complexity largely reflects high-dimensional covariance estimation difficulty rather than exploitable non-linearity. A simple L2-regularized logistic regression (0.952 AUROC) bounds or outperforms twelve controlled architectural alternatives, and our multi-layer aggregation exceeds CLAP cross-layer attention probing under matched paradigm. Because the signal spans a contiguous layer band, LayerMix aggregates it to match oracle-layer performance without oracle access. Our claims characterize the geometry within the controlled paired-example paradigm. Our code is available at https://github.com/js-lee-AI/LayerMix.

arXiv

Leveraging Turn-taking Dynamics for Intent Recognition in Multi-party Conversations

Galo Castillo-López, Alexis Lombard, Gaël De Chalendar, Nasredine Semmar

We propose a multi-task learning approach for multi-party dialogue intent recognition that leverages an auxiliary task that models turn-taking dynamics. Specifically, we introduce turn-transition entropy, a self-supervised target computed from the sequence of speaker transitions, which quantifies the predictability of interaction patterns. Experiments on multiple pre-trained models demonstrate that incorporating this auxiliary task improves intent recognition performance, outperforming existing approaches which ignore multi-party interaction dynamics. We find that our proposed continuous target can be learned as a single-task objective, suggesting that it is an actual signal carrying useful information.

arXiv

Causal Interventions Reveal Typologically Organized Syntactic Mechanisms in Multilingual Language Models

Sasha Boguraev, Toshiki Nakai, Kyle Mahowald, Julius Steuer

Linguistic theory has long recognized cross-linguistic syntactic regularities, leading to claims that these similar structures are processed by similar mechanisms. However, this hypothesis has been difficult to test empirically due to our lack of fine-grained, manipulable access of human processing mechanisms. In this work, we take advantage of techniques from mechanistic interpretability to study such a question in multilingual LMs. We first isolate language-internal mechanisms before attempting to transfer them cross-lingually. Across four models and three well-studied constructions (subject--verb number agreement, anaphoric pronoun gender agreement, and filler--gap object extraction) we find consistent cross-lingual mechanism transfer. We further find transfer to be graded, with more transfer between more typologically similar languages. We believe our work provides novel hypotheses about cross-linguistic syntactic structures and multilingual processing, and more broadly shows how the study of language models can help inform linguistic theory.

arXiv

VoiceCodeBench: Evaluating Exact Structured-Token Recovery in Automatic Speech Recognition

Tyler Baumgartner, Brandon Tai, Lisa Kaelin-Martin, Candice Fan, Luc Debaupte, Bill Wang, Yi Zhong

Automatic speech recognition (ASR) systems are commonly evaluated with word error rate (WER), yet many voice workflows depend on exact written values for identifiers, paths, and measured quantities. A transcript can appear fluent and achieve low WER while corrupting a value that a downstream system must parse, store, or execute. We introduce VoiceCodeBench, a benchmark for evaluating exact structured-token recovery in English ASR. It contains 300 human-recorded workplace segments spanning eight workflow domains and 1,482 audited target entities across 26 entity types, each with a canonical written form recoverable from the audio. Under a raw-audio-only protocol, systems receive audio bytes without additional context or metadata. Alongside WER, we evaluate Canonical Token/Entity Match (CTEM), Task Success Rate (TSR), and per-type exact recovery. Across 12 baseline ASR systems, lower WER generally corresponded to better structured-token recovery but did not fully determine it: Spearman correlations were -0.73 for both WER versus CTEM and WER versus TSR. The strongest baseline by TSR reached only 68.7%, leaving nearly one third of recordings with at least one unrecovered workflow-critical value. These results show that entity-sensitive metrics are needed to assess whether ASR output preserves exact values that production systems must parse, route, store, compare, or execute.

arXiv

SemKV: Semantic Mixed-Precision KV Cache Quantization Guided by the Quality Cliff for Long-Context LLM Inference

Daeha Lee, Do-Hyung Kim, Jae-Hong Kim

The key-value (KV) cache is the dominant memory bottleneck of long-context large language model (LLM) inference, growing linearly with context length. We show that uniform KV quantization on a fractional-bit grid does not degrade gracefully: under a prespecified multi-seed statistical protocol, Llama-3.1-8B-Instruct with an affine quantizer is statistically indistinguishable from FP16 KV down to 2.322 code bits/value and collapses at 2.0 bits - a quality cliff in (2.0, 2.322] that reappears in generation-time quantization and multi-turn dialogue and transfers to Mistral-7B. The cliff reframes importance-aware mixed precision: above it, eight model-internal importance indicators are statistically interchangeable, so the benefit of mixing is grid interpolation, reaching average precisions uniform quantization cannot realize. SemKV preserves every token, ranks tokens by a model-internal score, and assigns two adjacent above-cliff precisions, achieving a measured 6.0x storage reduction with no statistically detectable quality difference from full KV (n=900, three seeds), and outperforming FP16 token pruning granted a 1.5x larger memory budget. Replacing the affine base with a distortion-optimized quantizer (TurboQuant-MSE) lowers the cliff in every protocol tested, raising the no-detectable-loss operating point to 7.9x. The recipe: measure the cliff for the target deployment setting, then interpolate above it.

arXiv

Moving the Mean Toward the Known Good, Not Beyond It: What Inference-Time Interventions and Weight Consolidation Buy in Open-Ended Generation

Roberto I. Ono Filho

What does a generation loop gain from learning on its own verified successes? In cycles of generate, verify, select and LoRA-consolidate on online bin packing, training on value-filtered candidates shifts what the model writes on held-out variants toward value (-1.7 points of excess, p=0.008; -3.1 against a random-consolidation control, p=0.004) while the best observed candidate converges to the classic heuristic's level and no further. A confirmation battery replicates the whole procedure three times, with fresh seeds and a never-consulted held-out set read exactly once: the mean was nearly identical in all three lineages (-2.0, -1.8, -1.9), and after aggregating within held-out variant all seven evaluable variants favored consolidation (p=0.008). The best observed candidate moved to the classic heuristic's level, exactly (0.021028 in all three lineages, for attract and for the random control alike), and never beyond it. A matched SFT-only control shows the supervised anchor, not repulsion from bad candidates, does the concentrating (96% of candidates land exactly at the classic heuristic's level). The tails cut both ways: consolidation lowers the per-candidate rate of better-than-classic candidates (10% to 3.9%) while its larger production yields more such candidates absolutely (5 against 1, on few events). As motivation we report the inference-time ledger that led here: a model-written schematic recap buys judged document integration and nothing buys development; a verifier written into the stream is imitated, 16.4 fabricated verdict lines per notebook. Mean quality among valid candidates can be bought and replicated; the observed best goes to the classic and, so far, never beyond it.

arXiv

No Detectable Change in Side-Level WER from Prompt-Level Context: A Preregistered Ablation on a Production Oral-History Corpus

Theodore O. Cochran, Stephanie Dodson, Keith Nore

Supplying context at inference time to a large multimodal model is an inexpensive lever for adapting speech transcription to a domain, and earlier results on smaller models reported large gains. This work tested that mechanism where it ships, in the prompt-conditioning layer of a production oral-history transcription tool, on a sample from its own production corpus. Full prompt-level context did not detectably change side-level word error rate (WER), and none of the four preregistered hypotheses was supported. The design was a within-item paired ablation, preregistered with the analysis code frozen by hash before the confirmatory batch was scored; two disclosed gpt-4o pilot sides had been scored earlier, during scorer development. Nineteen cassette sides, about 10.6 hours of degraded 1970s-80s interview audio, were reprocessed through the production code path under three prompt arms, crossed with two deployed commercial configurations, gpt-4o-transcribe and gemini-2.5-flash, and scored against operator-corrected verbatim references. For gpt-4o-transcribe the median paired difference between the full-context and no-context arms was +0.6 WER points, with a side-resampled interval of [-1.1, +1.0]; the Gemini estimates were too unstable to support a comparable negative inference. A post-hoc rerun found run-to-run pipeline variability larger than the confirmatory differences, so effects of that size cannot be resolved from one transcription per cell. An implementation audit verified the manipulation was live, and sequence-alignment analysis found a small improvement on complete context-listed phrases, too small to materially change side-level WER, and for Gemini coexisting with worsened unlisted-token error. Evaluating context mechanisms therefore requires sequence-aligned term-level, insertion, and speaker-label measures alongside aggregate accuracy.

arXiv

Latent-Space Intervention for Cross-Lingual Factual Consistency: Consistency Improvements without Accuracy Drops

Faeze Ghorbanpour, Constanza Fierro, Alexander Fraser, Anders Sogaard

Large Language Models (LLMs) often answer the same factual question differently across languages. We study whether cross-lingual latent-space intervention can reduce this inconsistency. We train layer-specific autoencoders on parallel multilingual representations and apply inference-time corrections to factual QA prompts. We find that latent intervention improves geometric alignment between languages, and that this improvement translates into consistent gains in cross-lingual consistency with English across both open-ended and multiple-choice QA formats, without degrading factual accuracy. In open-ended QA, Spearman's rank correlation between English and non-English languages improves substantially, with gains of 0.16 for English-Arabic and 0.20 for English-Russian pairs. In multiple-choice QA, answer agreement with English improves consistently across both KLAR and mParaRel. Ablations show that AE reconstruction yields consistent gains at no accuracy cost, while PCA projection contributes marginally, and mean-shift produces substantially larger consistency gains in open-ended QA at the cost of some accuracy.

arXiv

The Halt Vector: Internalizing a Causal Steering Intervention for Efficient Reasoning

Dylan Jayabahu, Tinuade Adeleke

Reasoning models do not stop when they know the answer. On DeepSeek-R1-Distill-Qwen-7B the chain of thought runs about twice as long as the model's own answer probability takes to settle, and how much of that excess is removable varies from problem to problem, so a global length penalty cannot take it out. We take it out by internalizing a causal interpretability finding into the weights. The mechanism is a halt vector: a difference-of-means direction at layer 18 of this model whose steering strength controls how long it thinks, while a replicated value axis does nothing. Installing that intervention in the weights is harder than it looks. Maximizing the scalar projection onto the direction corrupts the off-axis dimensions a frozen downstream reader depends on, and generation gets longer instead of shorter; what works is reconstructing the whole steered activation with those dimensions pinned to their natural values. Fit from 24 problems and no reinforcement learning, the halt removes about a quarter of the thinking at held accuracy across five unseen benchmarks, and the cut tracks each problem's own removable slack at 0.70. It also closes a non-termination pathology that grows with difficulty and that a decoding-time confidence hook makes worse. We do not claim to beat a well-tuned length penalty or decoding-time early exit on the raw trade-off; the contribution is how the halt is obtained.

arXiv

A rigor-matched audit of periodic-step layer skipping for efficient llm inference: conflayers versus swift, with a supplemental analysis of trained routing alternatives

Prateek Kumar Sikdar, Arpan Ghosh

Layer-skipping methods for efficient LLM inference decide, at some granularity, which transformer layers to execute for a given input. We present a rigor-matched, three-seed audit of two periodic-step, search-based methods that make this decision online at inference time and re-evaluate it every few generation steps: a confidence-gated early-exit baseline (ConfLayers) and genuine self-speculative decoding (SWIFT, Xia et al. 2024), together with vanilla autoregressive decoding, across two model scales (Qwen2.5-0.5B and Qwen2.5-1.5B) and two tasks (GSM8K reasoning and CNN/DailyMail summarization). SWIFT is the strongest method on accuracy in three of four cells; ConfLayers is dominated everywhere, with particularly large deficits on GSM8K at 1.5B. Once online-search overhead is separated from pure inference cost, SWIFT's true inference speed is faster than ConfLayers's in all four cells (5-21%), reversing the naive wall-clock ranking in three of them. ConfLayers's search overhead is small and stable (1-2% of cost), while SWIFT's is larger and more variable (up to 28.7%). We additionally examine two trained-routing methods, LayerRoute (Sikdar, 2026) and LayerDrop (Fan et al. 2020), as a supplemental analysis because they operate at coarser decision granularities. Under a verified protocol with genuine per-input gating, a genuine full-model baseline, and genuine inference-time compute skipping, both show modest speedups (1.08-1.33x) but accuracy well below the periodic-step methods, including a near-total collapse for LayerRoute on GSM8K at 1.5B (0.003 mean exact-match across three seeds). We release the full audit protocol as a template for rigor-matched efficiency comparisons.

arXiv

Evaluating Multilingual Sentence Embeddings for Translation Error Detection:An English--Greek Contrastive Study

Eleftherios Kalogeros, Athanasios Ntalakas, Manolis Gergatsoulis, Paschalis Nikolaou, Sotiria-Lito Alexaki

Multilingual sentence embeddings are increasingly used to estimate semantic similarity across languages, yet their sensitivity to fine-grained translation errors remains insufficiently understood. This study investigates whether general-purpose multilingual embedding models can distinguish correct English-Greek translations from minimally modified erroneous alternatives. A contrastive dataset was developed from FLORES+ sentence-aligned reference translations and reviewed by two translation experts. It contains 1,850 examples across ten core and five exploratory error categories, covering factual, lexical-semantic, grammatical, relational, referential, and discourse-level phenomena. Five multilingual sentence-embedding models (BGE-M3, Multilingual E5, Multilingual MPNet, LaBSE, and Jina Embeddings v3) were evaluated using cosine similarity between each English source sentence and its correct and erroneous Greek translations. A reference-free COMETKiwi model was also evaluated as an MT quality-estimation baseline. Performance was assessed through contrastive accuracy and score margins for category-specific sensitivity. BGE-M3 achieved the highest accuracy among embedding models at 89.30 percent, while COMETKiwi achieved 94.49 percent. Embedding models detected explicit factual and lexical changes more reliably than tense-and-aspect and pronoun-coreference errors. COMETKiwi improved performance on several difficult categories, including tense and aspect, pronoun and coreference, and semantic-role errors, but showed lower sensitivity to date-and-time errors and underperformed the embedding models on numbers. The results show complementary error-sensitivity profiles: multilingual sentence embeddings provide useful semantic adequacy signals but are better suited as components of broader translation-evaluation frameworks than as standalone metrics.

arXiv

MultiHashFormer: Hash-based Generative Language Models

Huiyin Xue, Atsuki Yamaguchi, Nikolaos Aletras

Language models (LMs) represent tokens using embedding matrices that scale linearly with the vocabulary size. To constrain the parameter footprint, prior work proposes hashing many tokens into a single vector within encoder-only models. While this offers parameter efficiency, many-to-one collisions prevent its use in causal LMs. In this paper, we propose MultiHashFormer, a new framework that allows hash-based autoregression. Each token is represented as a unique hash signature, a short sequence of discrete hash IDs, generated by multiple independent hash functions. A Hash Encoder compresses this signature into a single latent vector for processing by a Transformer decoder. Then, a Hash Decoder generates the hash signature of the next token, which is then mapped back to text. We evaluate our approach at the 100M, 1B and 3B parameter scales, demonstrating that MultiHashFormer consistently outperforms standard Transformer LMs across multiple benchmarks. Furthermore, we show that our model handles multilingual vocabulary expansion with a constant parameter footprint without any modifications.

arXiv

DART: Draft-Agreement Routing for Training-Free Adaptive Thinking Budgets in Hybrid Reasoning Models

Jungseob Lee, Seongtae Hong, Seungjun Lee, Jaehyung Seo, Junyoung Son, Sugyeong Eo, Chanjun Park, Hyeongju Park, Hyeonseok Moon, Heuiseok Lim

Hybrid reasoning models can answer directly or spend extra tokens on extended thinking. A practical router should choose between these modes for each query, so easy problems avoid unnecessary reasoning and hard problems receive enough budget to finish the answer. Existing routers move in this direction, but they typically require labeled training data or fix thinking budgets up front, ignoring answer-level evidence from the model itself. We introduce DART, a training-free routing framework that samples two cheap no-think drafts, accepts direct answering when the drafts agree, and predicts a thinking budget from draft entropy when they disagree. Across the main comparisons, DART preserves or improves always-thinking accuracy in most settings while reducing thinking-token use. Accuracy improves by up to +9.0 points on Olympiad-level math and by up to +22.5 points on code under execution-based equivalence, while thinking-token use drops by 32-73%. The Stage~1 signal extends across model scales (0.6B--32B), model families, and API-only hosted settings, with no labeled data and no gradient updates required. Our code is available at https://github.com/js-lee-AI/DART.

arXiv

SciOrch: Learning to Orchestrate Expert LLMs for Solving Frontier Multimodal Scientific Reasoning Tasks

Jingru Guo, Xiangyuan Xue, Lian Zhang, Wanghan Xu, Siki Chen, Philip Torr, Wanli Ouyang, Lei Bai, Zhenfei Yin

Frontier scientific reasoning remains a major challenge for large language models (LLMs), where even the strongest commercial systems fall short of expert-level performance. A closer look at model behavior reveals substantial complementarity that single-model evaluation hides: different frontier models excel on different question types, and no single model captures the full picture. We present SciOrch, a framework that trains a lightweight 8B model to orchestrate frontier LLMs for scientific reasoning. The orchestrator decomposes each question, delegates sub-problems to selected commercial models through API calls, and synthesizes a final answer. Training such an orchestrator is fundamentally harder than conventional agentic RL: each action triggers an API call that is expensive in both dollar cost and latency, making standard online rollouts infeasible. We address this with MCTS-based approach, producing diverse orchestration trajectories, extracting per-node single-turn samples, and optimizing the orchestrator with GRPO-style training. On a 240-question test set spanning SGI-Reasoning and Scientists' First Exam, SciOrch reaches 56.66% average accuracy, outperforming the strongest single commercial model by 3.74% and the strongest multi-agent baseline by 3.33%. It also attains the best accuracy on both SGI and SFE with less than half the API cost of typical multi-agent methods.

继续浏览

更多 AI 顶会论文

30+ 个顶会，按会议、年份和主题浏览。

浏览全部会议

进入任一会议和年份，查看完整论文列表。

全部会议 →

搜索论文

按标题、作者或关键词，在 21 万+ 篇论文里搜索。

搜索论文 →