Models Detect Impossible Questions but Answer Anyway

Featured

01 Retrieval: Spend the Legal RAG Budget Wisely

Cloud deployments can call hosted models for quality. Local systems must respect data boundaries, hardware limits, and latency. The same legal assistant therefore needs two distinct engineering configurations.

This study built an Uzbek legal retrieval benchmark with 178 expert-labeled queries. Its end-to-end benchmark contains 504 expert-curated question-answer pairs. Reported results show only a modest gap between open and proprietary models.

Targeted retriever tuning narrowed that gap at relatively low cost. The team then trained UTE-1, an open Uzbek text embedding model. Long-context QA tuning demands far more hardware, while changing laws quickly make parameterized knowledge stale. A failed QLoRA experiment supports that choice.

The evaluation remains small. End-to-end scores also depend on model judges checked by humans and independent models. More work must test whether the findings transfer across languages and legal systems.

Key takeaways:

• Design cloud and local legal RAG deployments around their separate quality and data constraints.
• For resource-limited teams, targeted retriever tuning may beat long-context QA tuning on cost.
• Treat the updateable retrieval pipeline as a core asset instead of freezing legal knowledge into model parameters.

Source: Cloud and On-Premises Deployment of Uzbek Legal RAG via Targeted Retriever Fine-Tuning

02 Safety: Models Know When No Answer Exists

A model can answer a structurally impossible question such as cot(-540°) without missing the anomaly. Across instruction models from 1.7B to 70B parameters, hidden states contain a linear direction separating answerable and impossible problems.

That recognition direction is almost orthogonal to the refusal direction created by safety training. Intervening along it can move refusal behavior either way as intervention strength changes. Comparisons between base and instruction models suggest the disconnect already exists after pretraining.

Adding more refusal examples may not fix it. Developers should inspect how internal judgments reach the response policy instead of focusing only on knowledge and prompts. Current evidence covers structured math and coding problems, so broader tasks still need testing.

Key takeaways:

• When models answer impossible questions, first test whether they recognized the issue but failed to act.
• Safety refusal mechanisms may not consume signals the model already has about answerability.
• Evaluate anomaly recognition and final refusal separately to locate the routing failure between them.

Source: Recognition-Refusal Misalignment in LLMs: Why Models Answer Structurally Unanswerable Questions

03 Multimodal: CoVA-SFT Supervises Visual Reasoning States

CoVA-SFT turns explicit reasoning, agentic rendering, and verification loops into supervised multimodal steps. Models learn to alternate between language and visual abstractions during reasoning.

The dataset contains 51.9K samples and more than 222K multimodal reasoning steps. It addresses a weakness of text-only reasoning, which must serialize visual problems into words. Models can instead construct and revise visual abstractions during inference.

Fine-tuned models more than doubled the average CoVA-Bench score of other interleaved chain-of-thought baselines. They still trailed a strong text-only chain-of-thought baseline. Visual abstraction chains have not won outright.

Key takeaways:

• Mix language with visual abstractions instead of forcing every visual problem into text.
• Judge reasoning datasets by whether they supervise intermediate representations and correction loops.
• Doubling interleaved baselines is impressive, but trailing strong text reasoning calls for more validation.

Source: CoVA-SFT: A Large-Scale Dataset for Chain of Visual Abstractions