The model that critiques your question
Frontier models trust flawed premises, answer unanswerable questions, and lose track in long reasoning chains. Four preprints build the refusal layer.
- In a preprint, frontier LLMs are strongly premise-trusting by default and rarely critique flawed premises in text-to-SQL tasks; critique-oriented prompting improves detection and abstention but increases false invalid judgments, and the PremiseSQL benchmark contains 1,800 flawed queries across four premise-failure categories. Self-reported m:…9549
- In a preprint, LLMs show significant reliability gaps in rejecting unanswerable questions as context length scales; the FACTGUARD-BENCH dataset spans 25,220 examples with contexts up to 128K tokens, and training on it lifts unanswerable-question rejection accuracy from 71.62% to 81.17%. Self-reported m:…1601
- In a preprint testing reconciliatory explanation generation across 13 LLMs, longer reasoning chains are negatively associated with success at resolving contradictions, while reinforcement learning training on the task improves accuracy on the True Detective and MuSR narrative reasoning benchmarks. Self-reported m:…0030
- In a preprint on probe-based uncertainty estimation, structured and compressed features prove more robust than raw hidden states under distribution shift, and prompting and label construction significantly affect probe behavior. Self-reported m:…4661
Your query is broken; the model did not notice
In a preprint, the PremiseSQL authors test text-to-SQL systems on queries built on flawed premises and find that frontier LLMs are strongly premise-trusting by default, rarely critiquing the premise before generating executable SQL. The benchmark behind the finding contains 1,800 flawed natural-language-to-SQL queries across four categories of premise failures. A model that writes working SQL for a nonsense question is not being helpful; it is manufacturing data-integrity risk. m:…9549
Critique-oriented prompting improves detection and abstention on flawed queries, but it increases false invalid judgments on original, valid queries. There is no free refusal: every gain in skepticism costs something in false alarms. Product teams need to measure both sides of that trade, because a model that second-guesses correct queries will be turned off by its users. m:…9549
Unanswerable, at 128K
A second preprint takes the same problem into long context. The authors show that LLMs exhibit significant reliability gaps in rejecting unanswerable questions, and the gaps widen as context length scales. Their FACTGUARD-BENCH dataset contains 25,220 examples with context lengths up to 128K tokens, built with an evidence-grounded training workflow designed to teach models the difference between missing evidence and hallucinated answers. m:…1601
Training on the new benchmark lifts unanswerable-question rejection accuracy from 71.62% to 81.17%. That is real progress and still a long way from trustworthy: nearly one in five unanswerable questions still gets an answer. For anyone building RAG systems, the metric to watch is not just answer quality but refusal quality, and it degrades exactly where context gets longest. m:…1601
Contradictions break long chains
The third preprint introduces reconciliatory explanation generation: given apparently contradictory statements, can the model hypothesize an explanation that resolves them? Across 13 tested LLMs, the answer is mostly no, and longer reasoning chains are negatively associated with success. More thinking does not help; it gives the model more rope. m:…0030
Reinforcement learning training on reconciliatory explanation tasks improves accuracy on the True Detective and MuSR narrative reasoning benchmarks, which suggests the capability is trainable rather than architectural. The product implication, and this is reasoning, is that contradiction-handling should be an explicit training objective for any model deployed where inputs disagree, which is to say nearly everywhere. m:…0030
There is a subtle warning inside the chain-length finding. If longer reasoning correlates with worse reconciliation, then the fashion of ever-longer chains of thought may be actively degrading a capability enterprises need most: holding two conflicting claims in view and resolving them honestly. Reasoning depth and reasoning reliability are not the same axis, and optimizing the first can cost you the second. m:…0030
Uncertainty needs structure
The fourth preprint studies probe-based uncertainty estimation and finds that in-domain performance is a misleading metric: raw hidden states and attention features are difficult to beat in-domain but fail under distribution shift. Structured, compressed features prove more robust. Probes trained on the easy setting do not survive contact with the real one. m:…4661
Prompting and label construction significantly affect probe behavior, making them first-class variables in uncertainty estimation rather than implementation details. The thread connecting all four preprints, and this is reasoning, is that reliability is becoming a product surface: premise critique, abstention, contradiction reconciliation, calibrated uncertainty. The vendors that ship these as features, measured and priced, will win the enterprise deals. The ones selling raw capability will keep answering questions they should have refused. m:…4661
What is still unresolved
- What false-invalid rate enterprises will tolerate in exchange for premise critique
- Whether refusal quality can be maintained as context windows keep growing
Provenance by Synorb. Every claim links to its source. This story was written by an agent from the Manifests listed below. Read the methodology and limitations, or send a correction.
Free to reuse, including commercially. Credit “Hanging Context, with provenance by Synorb” and link to this story. Keep the cited Manifest ids. Reuse terms v1.