Abstract
Compact task-specialized language models demonstrate superior performance in multi-hop reasoning and faithfulness compared to larger general-purpose models through a novel training pipeline and structured reasoning traces.
Recent progress in the development of language models has been defined by scale, with each generation absorbing more of the world's knowledge into its weights. However, many practical applications benefit more from robust reasoning than from extensive parametric knowledge. In this setting, task-specialized small language models (SLMs) offer a principled design choice. We introduce Optimal Cognitive Core (OCC), a family of SLMs built around this premise. As a variant of OCC, we present OCC-RAG, optimized for faithful question answering (QA) grounded in the provided context. This task directly aligns with the OCC design approach, requiring multi-hop reasoning over supplied passages while ignoring memorized knowledge. To train OCC-RAG, we implement a novel pipeline for synthesizing multi-context, multi-hop QA data at scale, producing a corpus of over three million examples targeting multi-hop reasoning, strict context faithfulness, and calibrated abstention. We release OCC-RAG-0.6B and OCC-RAG-1.7B, both mid-trained on this corpus. The models produce structured reasoning traces with source citations grounded in literal quotes from the context. Through OCC-RAG, we demonstrate that compact, task-specialized SLMs can match or exceed general-purpose models 2 -- 6x their size across multi-hop reasoning (HotpotQA, MuSiQue, TAT-QA), faithfulness (ConFiQA), and refusal (MuSiQue-Un) benchmarks.
Community
Recent progress in the development of language models has been defined by scale, with each generation absorbing more of the world's knowledge into its weights. However, many practical applications benefit more from robust reasoning than from extensive parametric knowledge. In this setting, task-specialized small language models (SLMs) offer a principled design choice. We introduce Optimal Cognitive Core (OCC), a family of SLMs built around this premise. As a variant of OCC, we present OCC-RAG, optimized for faithful question answering (QA) grounded in the provided context. This task directly aligns with the OCC design approach, requiring multi-hop reasoning over supplied passages while ignoring memorized knowledge. To train OCC-RAG, we implement a novel pipeline for synthesizing multi-context, multi-hop QA data at scale, producing a corpus of over three million examples targeting multi-hop reasoning, strict context faithfulness, and calibrated abstention. We release OCC-RAG-0.6B and OCC-RAG-1.7B, both mid-trained on this corpus. The models produce structured reasoning traces with source citations grounded in literal quotes from the context. Through OCC-RAG, we demonstrate that compact, task-specialized SLMs can match or exceed general-purpose models 2 - 6x their size across multi-hop reasoning (HotpotQA, MuSiQue, TAT-QA), faithfulness (ConFiQA), and refusal (MuSiQue-Un) benchmarks.
Congratulations on this release — the focus on faithful, context-grounded QA with calibrated abstention and citation-anchored reasoning traces is really nice work, and getting 0.6B/1.7B models to match or beat much larger ones on ConFiQA faithfulness is an impressive result. We especially liked the explicit "query analysis → source analysis → reasoning → status → answer" structure and the not enough information abstention behavior.
We've been thinking along very similar lines and wanted to share our work, in case it's of interest:
CANOE: Teaching Large Language Models to Maintain Contextual Faithfulness via Synthetic Tasks and Reinforcement Learning (AAAI 2026, Oral)
Code: https://github.com/S1s-Z/CANOE
Paper: https://arxiv.org/abs/2505.16483
Like OCC-RAG, CANOE improves contextual faithfulness without human annotation by synthesizing easily-verifiable short-form QA data across diverse tasks. The main difference is the post-training recipe: we propose Dual-GRPO, a rule-based RL method with tailored rewards that jointly optimizes both short-form and long-form generation (avoiding the over-optimization you get from short-form data alone, and removing the need to train reward models on labeled preference data). It improves faithfulness across 11 downstream tasks.
There seems to be a lot of shared ground between the two approaches (synthetic data for faithfulness, abstention/grounding), and possibly complementary ideas (your citation-anchored traces + small-model specialization vs. our RL-based optimization). We'd be glad if you took a look, and we'll be following your work going forward. Happy to exchange notes anytime.
Get this paper in your agent:
hf papers read 2606.00683 Don't have the latest CLI?
curl -LsSf https://hf.co/cli/install.sh | bash Models citing this paper 2
Datasets citing this paper 0
No dataset linking this paper
Spaces citing this paper 0
No Space linking this paper
Collections including this paper 0
No Collection including this paper