[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"article-teaching-llms-when-to-trust-context-en":3,"article-related-teaching-llms-when-to-trust-context-en":29,"series-research-cf660b1b-17fc-47e6-ad3a-360fb922ca9e":75},{"id":4,"slug":5,"title":6,"content":7,"summary":8,"source":9,"source_url":10,"author":11,"image_url":12,"cover_image":12,"category":13,"language":14,"translated_content":11,"related_article_id":15,"keywords":16,"key_takeaways":22,"views":26,"created_at":27,"published_at":28,"topic_cluster_id":11},"cf660b1b-17fc-47e6-ad3a-360fb922ca9e","teaching-llms-when-to-trust-context-en","Teaching LLMs When to Trust Context","\u003Cp data-speakable=\"summary\">MIST and SCOPE reduce context-induced errors while keeping models useful when context is trustworthy.\u003C\u002Fp>\u003Cul>\u003Cli>\u003Cstrong>Research org\u003C\u002Fstrong>: Unspecified in arXiv abstract\u003C\u002Fli>\u003Cli>\u003Cstrong>Core data\u003C\u002Fstrong>: Four matched conditions per reasoning item\u003C\u002Fli>\u003Cli>\u003Cstrong>Breakthrough\u003C\u002Fstrong>: Selective trust training with balanced DPO preference pairs\u003C\u002Fli>\u003C\u002Ful>\u003Cp>Large language models are increasingly expected to use outside context: retrieved passages, tool outputs, user-provided notes, or other signals that may or may not be reliable. The catch is simple but painful: one misleading context snippet can turn a correct answer into a wrong one. This paper argues that the real problem is not just “how do we make models robust to bad context?” but “how do we make them trust context selectively?”\u003C\u002Fp>\u003Cp>That distinction matters for anyone building retrieval-augmented systems, assistants that read documents, or agents that combine model knowledge with external evidence. A model that ignores context altogether may look safe in a narrow test, but it will also fail when the context is actually correct and useful. The paper’s main claim is that we should measure and train for selective trust, not blind resistance.\u003C\u002Fp>\u003Ch2>What problem this paper is trying to fix\u003C\u002Fh2>\u003Cp>The abstract points to a failure mode that is easy to miss in ordinary evaluation. If you train or test a model only on misleading context, you can end up rewarding a model that simply learns to distrust everything. That sounds robust, but it creates a different problem: the model becomes less capable of using correct context when it should.\u003C\u002Fp>\n\u003Cfigure class=\"my-6\">\u003Cimg src=\"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1786082582487-m31g.png\" alt=\"Teaching LLMs When to Trust Context\" class=\"rounded-xl w-full\" loading=\"lazy\" \u002F>\u003C\u002Ffigure>\n\u003Cp>In practical terms, that is a bad trade for developers. Many production systems are built on the assumption that external context is usually helpful, but not always. A support bot may receive a corrupted snippet. A coding assistant may see an outdated \u003Ca href=\"\u002Ftag\u002Fapi\">API\u003C\u002Fa> example. A document QA system may retrieve a passage that is irrelevant or subtly wrong. The model has to decide when to lean on the context and when to ignore it.\u003C\u002Fp>\u003Cp>According to the abstract, the authors treat this as a selective trust problem. They also report that susceptibility to misleading signals is universal across a comprehensive \u003Ca href=\"\u002Ftag\u002Fbenchmark\">benchmark\u003C\u002Fa> study, which suggests this is not a niche quirk of one model family or one dataset. The abstract does not give benchmark numbers for that study beyond the new metric design, so there is no specific accuracy table to quote here.\u003C\u002Fp>\u003Ch2>How MIST and SC2W work\u003C\u002Fh2>\u003Cp>The first contribution is MIST, a human-annotated benchmark built around matched conditions for each reasoning item. Each item is rendered in four versions: clean, misleading, correct-context, and irrelevant-context. That setup is important because it lets the authors compare how the same underlying question behaves when the surrounding signal changes.\u003C\u002Fp>\u003Cp>From an engineering perspective, that is a much better test than a single “bad context” dataset. It separates several cases that often get blurred together: context that helps, context that hurts, context that is irrelevant, and context that is simply absent or clean. If you only test one of those, you do not really know whether the model is learning selective use of evidence or just learning a blanket heuristic.\u003C\u002Fp>\u003Cp>The second contribution is SC2W, a paired metric that counts how often a misleading signal flips a clean-correct answer to wrong. In other words, it measures susceptibility directly. The abstract does not provide a numeric SC2W score in the summary text, but it does make clear that this metric is designed to capture the specific failure case the paper cares about.\u003C\u002Fp>\u003Cul>\u003Cli>MIST gives each reasoning item four matched context conditions.\u003C\u002Fli>\u003Cli>SC2W measures flips from clean-correct to wrong under misleading context.\u003C\u002Fli>\u003Cli>The evaluation setup is built to separate trust from blanket resistance.\u003C\u002Fli>\u003C\u002Ful>\u003Ch2>What SCOPE changes in training\u003C\u002Fh2>\u003Cp>The training method is called SCOPE, short for Selective Context Preference Optimization. The abstract says it mines clean-correct and misleading-wrong failures, then optimizes a standard Direct Preference Optimization objective over matched preference pairs. The key detail is that the pairs are balanced equally across all four conditions, instead of focusing only on misleading items.\u003C\u002Fp>\n\u003Cfigure class=\"my-6\">\u003Cimg src=\"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1786082576347-bb91.png\" alt=\"Teaching LLMs When to Trust Context\" class=\"rounded-xl w-full\" loading=\"lazy\" \u002F>\u003C\u002Ffigure>\n\u003Cp>That design choice is the heart of the paper. If you optimize only on misleading cases, you may teach the model to reject context too aggressively. By balancing across clean, correct-context, misleading, and irrelevant-context conditions, SCOPE tries to preserve the ability to benefit from good context while still reducing the harm from bad context.\u003C\u002Fp>\u003Cp>For developers, this is the part that will feel most familiar. It is still DPO at the core, which means the method fits into an existing preference-optimization pipeline rather than requiring a brand-new training stack. The novelty is not a new model architecture; it is the way the preference pairs are selected and balanced around context trust behavior.\u003C\u002Fp>\u003Ch2>What the paper actually shows\u003C\u002Fh2>\u003Cp>The abstract says the authors ran a comprehensive benchmark study and found that susceptibility to misleading signals is universal. It also says SCOPE substantially reduces SC2W on popular open-sourced models while preserving accuracy when the added context is clean, correct, or irrelevant.\u003C\u002Fp>\u003Cp>That is the practical headline: lower flip rates without sacrificing performance on useful or harmless context. The abstract does not include the exact reduction percentages or model-by-model benchmark numbers, so this article should not pretend otherwise. What we can say is that the authors claim a consistent improvement on the selective-trust metric, not just a narrow win on one dataset.\u003C\u002Fp>\u003Cp>This is an important nuance. In many context-robustness papers, the model gets better at ignoring noise but worse at using signal. Here the stated goal is to avoid that collapse. The paper is explicitly arguing that the right success criterion is not “can the model ignore misleading context?” but “can it ignore misleading context without discarding trustworthy context too?”\u003C\u002Fp>\u003Ch2>Why developers should care\u003C\u002Fh2>\u003Cp>If you are building systems that depend on retrieval, tool outputs, or user-supplied context, this paper maps to a real production risk. A model that over-trusts context can be manipulated by a bad retrieval result. A model that under-trusts context can become brittle, unhelpful, or blind to the very evidence you added to improve it.\u003C\u002Fp>\u003Cp>That means evaluation needs to be more precise than a single accuracy number. You want to know how often bad context flips a correct answer, but also whether good context still helps. MIST and SC2W give the paper a way to ask both questions at once.\u003C\u002Fp>\u003Cp>The limitation, based on the abstract alone, is that we do not get the full experimental detail here: no exact benchmark scores, no per-task breakdown, and no explicit statement about which open-source models were tested. We also do not know from the abstract how much human annotation effort MIST required or how broadly it generalizes beyond the benchmark setting.\u003C\u002Fp>\u003Cp>Even with those gaps, the engineering message is clear. Context-aware \u003Ca href=\"\u002Ftag\u002Fllms\">LLMs\u003C\u002Fa> should not be judged only by their resistance to bad inputs. They should be judged by whether they can selectively trust context when it is helpful, ignore it when it is misleading, and stay stable when it is irrelevant.\u003C\u002Fp>\u003Ch2>Bottom line\u003C\u002Fh2>\u003Cp>SCOPE is a training recipe for selective trust, and MIST is the benchmark that makes that behavior measurable. Together, they push the field away from a simplistic “ignore bad context” mindset and toward a more useful standard: use context when it deserves trust, and reject it when it does not.\u003C\u002Fp>\u003Cp>For teams shipping retrieval-augmented or tool-using assistants, that is the right problem to solve. Robustness is not the same as refusal. The paper’s core argument is that good systems need both skepticism and judgment.\u003C\u002Fp>","MIST and SCOPE reduce context-induced errors while keeping models useful when context is trustworthy.","arxiv.org","https:\u002F\u002Farxiv.org\u002Fabs\u002F2608.06377",null,"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1786082582487-m31g.png","research","en","66a9a233-e2fe-4726-a3e7-f01a6b47d100",[17,18,19,20,21],"LLMs","context robustness","DPO","retrieval-augmented generation","benchmark",[23,24,25],"Selective trust is a better target than blanket resistance to context.","MIST evaluates four matched context conditions per reasoning item.","SCOPE uses balanced DPO pairs to reduce misleading-context flips.",2,"2026-08-07T06:02:33.147809+00:00","2026-08-07T06:02:33.122+00:00",{"tags":30,"relatedLang":34,"relatedPosts":38},[31,32],{"name":21,"slug":21},{"name":17,"slug":33},"llms",{"id":15,"slug":35,"title":36,"language":37},"teaching-llms-when-to-trust-context-zh","教 LLM 何時信上下文","zh",[39,45,51,57,63,69],{"id":40,"slug":41,"title":42,"cover_image":43,"image_url":43,"created_at":44,"category":13},"7368755d-86ca-461e-9d95-d7e74e95b561","astra-turns-long-math-tasks-into-multi-agent-work-en","Astra turns long math tasks into multi-agent work","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1786088006851-29p0.png","2026-08-07T07:32:52.200646+00:00",{"id":46,"slug":47,"title":48,"cover_image":49,"image_url":49,"created_at":50,"category":13},"e69199db-e1f8-4e12-aaf2-ea92eeb2e0cc","evidence-linked-feature-engineering-heart-failure-en","Evidence-linked feature engineering for heart failure","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1786086190016-bykl.png","2026-08-07T07:02:31.382531+00:00",{"id":52,"slug":53,"title":54,"cover_image":55,"image_url":55,"created_at":56,"category":13},"4082af89-fbca-47cf-885c-52f6a90e6bfd","tool-calling-as-code-bfcl-v4-en","Why tool calling may work better as code","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1786084371891-2a3z.png","2026-08-07T06:32:26.114261+00:00",{"id":58,"slug":59,"title":60,"cover_image":61,"image_url":61,"created_at":62,"category":13},"fa5321dc-f874-4ab7-9a8c-427e02701561","cuda-binaries-turn-ptx-into-elf-you-can-inspect-en","CUDA binaries turn PTX into ELF you can inspect","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1786068197247-0rmx.png","2026-08-07T02:02:54.048314+00:00",{"id":64,"slug":65,"title":66,"cover_image":67,"image_url":67,"created_at":68,"category":13},"53a5becc-78b2-4c0a-b794-5c45df5bcf41","octolong-cross-repository-code-contexts-en","OctoLong trains LMs on cross-repo code context","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1785999775951-z3au.png","2026-08-06T07:02:27.420665+00:00",{"id":70,"slug":71,"title":72,"cover_image":73,"image_url":73,"created_at":74,"category":13},"011efa03-c54d-430a-8810-01f0a5422ad2","argus-self-evolving-runtime-long-tasks-en","Argus: a self-evolving runtime for long tasks","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1785997979079-uj2z.png","2026-08-06T06:32:35.693028+00:00",[76,81,86,91,96,101,106,111,116,121],{"id":77,"slug":78,"title":79,"created_at":80},"a2715e72-1fe8-41b3-abb1-d0cf1f710189","ai-predictions-2026-big-changes-en","AI Predictions for 2026: Brace for Big Changes","2026-03-26T01:25:07.788356+00:00",{"id":82,"slug":83,"title":84,"created_at":85},"8404bd7b-4c2f-4109-9ec4-baf29d88af2b","ml-papers-of-the-week-github-research-desk-en","ML Papers of the Week Turns GitHub Into a Research Desk","2026-03-27T01:11:39.480259+00:00",{"id":87,"slug":88,"title":89,"created_at":90},"87897a94-8065-4464-a016-1f23e89e17cc","ai-ml-conferences-to-watch-in-2026-en","AI\u002FML Conferences to Watch in 2026","2026-03-27T01:51:54.184108+00:00",{"id":92,"slug":93,"title":94,"created_at":95},"6f1987cf-25f3-47a4-b3e6-db0997695be8","openclaw-agents-manipulated-self-sabotage-en","OpenClaw Agents Can Be Manipulated Into Failure","2026-03-28T03:03:18.899465+00:00",{"id":97,"slug":98,"title":99,"created_at":100},"a53571ad-735a-4178-9f93-cb09b699d99c","vega-driving-language-instructions-en","Vega: Driving with Natural Language Instructions","2026-03-28T14:54:04.698882+00:00",{"id":102,"slug":103,"title":104,"created_at":105},"a34581d6-f36e-46da-88bb-582fb3e7425c","personalizing-autonomous-driving-styles-en","Drive My Way: Personalizing Autonomous Driving Styles","2026-03-28T14:54:26.148181+00:00",{"id":107,"slug":108,"title":109,"created_at":110},"2bc1ad7f-26ce-4f02-9885-803b35fd229d","training-knowledge-bases-writeback-rag-en","Training Knowledge Bases with WriteBack-RAG","2026-03-28T14:54:45.643433+00:00",{"id":112,"slug":113,"title":114,"created_at":115},"71adc507-3c54-4605-bbe2-c966acd6187e","packforcing-long-video-generation-en","PackForcing: Efficient Long-Video Generation Method","2026-03-28T14:55:02.646943+00:00",{"id":117,"slug":118,"title":119,"created_at":120},"675942ef-b9ec-4c5f-a997-381250b6eacb","pixelsmile-facial-expression-editing-en","PixelSmile Framework Enhances Facial Expression Editing","2026-03-28T14:55:20.633463+00:00",{"id":122,"slug":123,"title":124,"created_at":125},"6954fa2b-8b66-4839-884b-e46f89fa1bc3","adaptive-block-scaled-data-types-en","IF4: Smarter 4-Bit Quantization That Adapts to Your Data","2026-03-31T06:00:36.65963+00:00"]