# Question Answering

Federated learning usually requires institutions to share one model backbone so parameter updates can be aggregated, which keeps centers running different model families or sizes from collaborating. FedVP moves collaboration from parameter aggregation to shared verification: each client trains a verifier on its local data and candidate responses from its own frozen LLM, and the server pools these independently trained verifiers to score and select answers, never merging backbone parameters. Across five medical question-answering datasets, with centers using Qwen and LLaMA backbones from 3B to 14B parameters, one verifier pool is reused for all five backbones without retraining and raises mean accuracy by 1.46 to 5.79 percentage points over the strongest conventional federated baselines.

Federated learning usually requires institutions to share one model backbone so parameter updates can be aggregated, which keeps centers running different model families or sizes from collaborating. FedVP moves collaboration from parameter aggregation to shared verification: each client trains a verifier on its local data and candidate responses from its own frozen LLM, and the server pools these independently trained verifiers to score and select answers, never merging backbone parameters. Across five medical question-answering datasets, with centers using Qwen and LLaMA backbones from 3B to 14B parameters, one verifier pool is reused for all five backbones without retraining and raises mean accuracy by 1.46 to 5.79 percentage points over the strongest conventional federated baselines.

Geopolitical intelligence QA now has a public benchmark of 415 reports and 5,525 question-answer pairs, where document-level retrieval is nearly saturated but even with gold evidence 70% of risk-oriented answers fall short.

Geopolitical intelligence QA now has a public benchmark of 415 reports and 5,525 question-answer pairs, where document-level retrieval is nearly saturated but even with gold evidence 70% of risk-oriented answers fall short.

Multi-turn LLM agent training now ties exploration and credit assignment to one hindsight score: turns that lower the reference-answer likelihood are revisited and re-branched from the same parent, and inverse weights fold the outcomes back into the original policy, giving the highest aggregate exact match across seven QA benchmarks and three Qwen backbones.

Multi-turn LLM agent training now ties exploration and credit assignment to one hindsight score: turns that lower the reference-answer likelihood are revisited and re-branched from the same parent, and inverse weights fold the outcomes back into the original policy, giving the highest aggregate exact match across seven QA benchmarks and three Qwen backbones.

Open-domain multimodal document QA requires first locating the right passages and then reading them, yet current retrieval-augmented agents give one model both kinds of tools and let it schedule itself. KnowForge separates the pipeline by job: a search-only Locator agent, a code-scheduled bridge judge that forces the missing second hop, a bounded evidence pool merging every search ranking, and a Deep Reader that only reads text, images and tables. With a self-hosted Qwen3.6-35B-A3B, it reaches F1 47.5 on the full M3DocVQA (2,441 questions, 3,368 PDFs), 11.0 above the published M3DocRAG and 7.1 above one-shot retrieve-and-answer on the same index and model.

Open-domain multimodal document QA requires first locating the right passages and then reading them, yet current retrieval-augmented agents give one model both kinds of tools and let it schedule itself. KnowForge separates the pipeline by job: a search-only Locator agent, a code-scheduled bridge judge that forces the missing second hop, a bounded evidence pool merging every search ranking, and a Deep Reader that only reads text, images and tables. With a self-hosted Qwen3.6-35B-A3B, it reaches F1 47.5 on the full M3DocVQA (2,441 questions, 3,368 PDFs), 11.0 above the published M3DocRAG and 7.1 above one-shot retrieve-and-answer on the same index and model.

RAG-Stress introduces a controlled diagnostic protocol that holds the question and reference answer fixed, edits one assertion so the evidence supports a designated incorrect answer, and crosses two source-priority policies with three positions of the answer span in the evidence text, measuring a misleading rate on questions each model answered correctly without retrieval across fifteen systems, three QA datasets, and English and Chinese medical QA. Instructions that prioritize documents consistently produce higher misleading rates than those permitting prior knowledge, with a mean gap of 10.9–13.5 percentage points across the three QA datasets, and model-averaged rates follow End, Beginning, Middle under both policies.

RAG-Stress introduces a controlled diagnostic protocol that holds the question and reference answer fixed, edits one assertion so the evidence supports a designated incorrect answer, and crosses two source-priority policies with three positions of the answer span in the evidence text, measuring a misleading rate on questions each model answered correctly without retrieval across fifteen systems, three QA datasets, and English and Chinese medical QA. Instructions that prioritize documents consistently produce higher misleading rates than those permitting prior knowledge, with a mean gap of 10.9–13.5 percentage points across the three QA datasets, and model-averaged rates follow End, Beginning, Middle under both policies.

RECAST casts evidence construction as a sequential decision process over retrieval and computation operations: a lightweight RouterLM iteratively selects lexical, semantic, or relational primitives, or requests customized operations that a frozen CompilerLM translates into executable code, then passes accepted evidence to a frozen AnswerLM; RouterLM is trained with SFT followed by GRPO, achieving the strongest average success rate across six heterogeneous benchmarks and zero-shot generalization on three held-out benchmarks.

RECAST casts evidence construction as a sequential decision process over retrieval and computation operations: a lightweight RouterLM iteratively selects lexical, semantic, or relational primitives, or requests customized operations that a frozen CompilerLM translates into executable code, then passes accepted evidence to a frozen AnswerLM; RouterLM is trained with SFT followed by GRPO, achieving the strongest average success rate across six heterogeneous benchmarks and zero-shot generalization on three held-out benchmarks.