[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"article-test-time-harnesses-weak-model-transfer-en":3,"article-related-test-time-harnesses-weak-model-transfer-en":29,"series-research-5a953549-e09c-43e6-856c-63c394e85997":73},{"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},"5a953549-e09c-43e6-856c-63c394e85997","test-time-harnesses-weak-model-transfer-en","Test-Time Harnesses Transfer Skills Without Retraining","\u003Cp data-speakable=\"summary\">Strong models can build test-time harnesses that nearly double weaker models’ performance without updating parameters.\u003C\u002Fp>\u003Cul>\u003Cli>\u003Cstrong>Research org\u003C\u002Fstrong>: Unspecified in arXiv abstract\u003C\u002Fli>\u003Cli>\u003Cstrong>Core data\u003C\u002Fstrong>: Average target-model performance rose from 0.49 to 0.91\u003C\u002Fli>\u003Cli>\u003Cstrong>Breakthrough\u003C\u002Fstrong>: Stronger builder models iteratively construct inference-time harnesses\u003C\u002Fli>\u003C\u002Ful>\u003Cp>Until now, most capability transfer has meant retraining the smaller model. This paper asks a different question: can a stronger model help a weaker one at test time, with no parameter updates at all?\u003C\u002Fp>\u003Cp>That matters because retraining is expensive, slow, and often impractical once a model is already deployed. If the answer is yes, engineers get a new tool: a way to improve reliability by changing the \u003Ca href=\"\u002Ftag\u002Finference\">inference\u003C\u002Fa> setup instead of touching weights.\u003C\u002Fp>\u003Ch2>What problem this paper is trying to fix\u003C\u002Fh2>\u003Cp>The paper starts from a familiar distillation pattern. In conventional setups, a large teacher transfers skill to a smaller student through training-time methods such as teacher forcing or on-policy distillation. Those approaches work, but they require updates to the target model’s parameters.\u003C\u002Fp>\n\u003Cfigure class=\"my-6\">\u003Cimg src=\"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1786604571709-b71l.png\" alt=\"Test-Time Harnesses Transfer Skills Without Retraining\" class=\"rounded-xl w-full\" loading=\"lazy\" \u002F>\u003C\u002Ffigure>\n\u003Cp>Here, the authors explore whether transfer can happen after training is done. The target model stays frozen. Instead, a stronger builder model creates a harness around the weaker model so it can answer more reliably during inference.\u003C\u002Fp>\u003Cp>The key idea is simple: if the model itself is unstable, maybe part of the problem can be moved outside the model. The harness can take over some of the reasoning structure, route inputs differently, or constrain outputs more tightly.\u003C\u002Fp>\u003Ch2>How the method works in plain English\u003C\u002Fh2>\u003Cp>The paper calls this \u003Cem>strong-to-weak scaffolding\u003C\u002Fem>. A stronger builder model constructs inference-time harnesses for a weaker target model, and those harnesses are refined iteratively over multiple rounds.\u003C\u002Fp>\u003Cp>In the reported setup, each builder model uses 5% of the data as a validation set to tune the harness. After that iterative refinement, the finalized harness is evaluated on the full test set. No parameter updates are made to the target model.\u003C\u002Fp>\u003Cp>The authors study this on four representative Theory-of-Mind benchmarks. The abstract does not list the \u003Ca href=\"\u002Ftag\u002Fbenchmark\">benchmark\u003C\u002Fa> names, so there is no way to compare task-by-task results from the source alone. What it does make clear is that the method is tested across multiple benchmarks rather than on a single narrow case.\u003C\u002Fp>\u003Cp>What counts as a harness here? Based on the abstract, it is not a learned adapter or a fine-tuned prompt alone. It is a test-time structure that can include deterministic code, benchmark-specific routing, and strict answer-format enforcement. In other words, the harness is doing some of the work that the weaker model struggles to do consistently on its own.\u003C\u002Fp>\u003Ch2>What the paper actually shows\u003C\u002Fh2>\u003Cp>The headline result is large: average target-model performance rises from 0.49 to 0.91. That is close to a doubling, and it happens without retraining the target model.\u003C\u002Fp>\n\u003Cfigure class=\"my-6\">\u003Cimg src=\"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1786604573444-vcxd.png\" alt=\"Test-Time Harnesses Transfer Skills Without Retraining\" class=\"rounded-xl w-full\" loading=\"lazy\" \u002F>\u003C\u002Ffigure>\n\u003Cp>Just as important, the paper tries to explain where the gain comes from. The gains are mainly attributed to offloading unstable model reasoning into deterministic code, adding benchmark-specific routing, and enforcing strict answer formats. The paper says the improvement does \u003Cem>not\u003C\u002Fem> mainly come from making the target model reason more extensively or sample more broadly.\u003C\u002Fp>\u003Cp>The authors also report three additional patterns. First, builder-model reasoning effort improves harness quality monotonically. Second, platform effects are modest compared with the builder model’s own capability. Third, weaker target models get the largest gains.\u003C\u002Fp>\u003Cp>That last point is especially relevant for deployment. If the weakest models benefit the most, then harnesses may be most useful where teams care about cost, latency, or local deployment and cannot simply swap in a larger model.\u003C\u002Fp>\u003Ch2>Why developers should care\u003C\u002Fh2>\u003Cp>This paper is interesting because it reframes optimization at inference time. Instead of asking, “How do we make the model smarter through training?” it asks, “How do we structure the surrounding system so the model behaves better right now?”\u003C\u002Fp>\u003Cp>For developers, that opens up a practical design space. A frozen model can still be paired with logic that validates inputs, routes cases, constrains outputs, or delegates brittle reasoning to deterministic code. That means some reliability gains may be possible without a training pipeline.\u003C\u002Fp>\u003Cp>It also suggests a new complement to distillation. Training-time compression is still valuable, but it may not be the only way a strong model can help a weaker one. In this framing, the strong model can transfer not just outputs, but cognitive structure.\u003C\u002Fp>\u003Ch2>Limitations and open questions\u003C\u002Fh2>\u003Cp>The abstract gives strong directional results, but it leaves important details out. We do not get the names of the four benchmarks, the exact harness designs, or per-\u003Ca href=\"\u002Fnews\u002Fbenchmark-scores-dont-predict-your-bill-en\">benchmark scores\u003C\u002Fa>. We also do not get runtime cost, latency impact, or how much manual engineering the harnesses required.\u003C\u002Fp>\u003Cp>Because the method depends on a stronger builder model, the quality of the transfer is still tied to the builder’s own reasoning ability. The paper says builder reasoning effort helps and platform effects are modest, but that still leaves open how portable the approach is across model families and deployment environments.\u003C\u002Fp>\u003Cp>There is also a scope question. The paper focuses on Theory-of-Mind benchmarks, so the results may not generalize to every task class. The abstract does not claim broader benchmark coverage, and it does not provide evidence for domains outside the reported setting.\u003C\u002Fp>\u003Cp>Even with those limits, the core message is useful: some of what we think of as “model intelligence” can be shifted into the inference layer. For teams building AI systems, that is a reminder to treat harness design as a first-class engineering problem, not just a wrapper around the model.\u003C\u002Fp>\u003Ch2>Bottom line\u003C\u002Fh2>\u003Cp>This paper shows that a strong model can build a test-time harness that materially boosts a weaker model without changing its weights. The result is a practical alternative to retraining when you want more reliable outputs from a frozen model.\u003C\u002Fp>\u003Cp>For engineers, the takeaway is not that harnesses replace training. It is that inference-time scaffolding can be a powerful lever, especially when the target model is weak and the task benefits from deterministic structure.\u003C\u002Fp>\u003Cul>\u003Cli>Test-time scaffolding can transfer capability without parameter updates.\u003C\u002Fli>\u003Cli>Deterministic code and strict formatting appear to drive much of the gain.\u003C\u002Fli>\u003Cli>The approach looks most helpful for weaker target models, but task coverage is still narrow.\u003C\u002Fli>\u003C\u002Ful>","Strong models can build test-time harnesses that nearly double weaker models’ performance without updating parameters.","arxiv.org","https:\u002F\u002Farxiv.org\u002Fabs\u002F2608.12307",null,"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1786604571709-b71l.png","research","en","a212bb55-ce9d-4c3e-870d-8d6cd0ff3b76",[17,18,19,20,21],"test-time transfer","harnesses","distillation","theory of mind","inference-time scaffolding",[23,24,25],"Strong models can improve weaker ones at test time without retraining.","Most of the gain comes from deterministic harness logic, not more model reasoning.","The paper is promising but limited to four Theory-of-Mind benchmarks in the abstract.",1,"2026-08-13T07:02:25.968754+00:00","2026-08-13T07:02:25.956+00:00",{"tags":30,"relatedLang":32,"relatedPosts":36},[31],{"name":19,"slug":19},{"id":15,"slug":33,"title":34,"language":35},"test-time-harnesses-weak-model-transfer-zh","測試時外掛讓弱模型升級","zh",[37,43,49,55,61,67],{"id":38,"slug":39,"title":40,"cover_image":41,"image_url":41,"created_at":42,"category":13},"84526e03-6caf-4b7e-a910-64f2b712da66","dreamfly-aerial-vln-memory-planning-en","DreamFly improves aerial VLN with memory and planning","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1786602779859-brjh.png","2026-08-13T06:32:24.41046+00:00",{"id":44,"slug":45,"title":46,"cover_image":47,"image_url":47,"created_at":48,"category":13},"84e73ffd-ac52-4e86-88cc-abb15eeb9c5e","ava-encoder-agent-native-video-representation-en","AVA-Encoder turns films into editable knowledge graphs","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1786600972460-6yms.png","2026-08-13T06:02:22.174937+00:00",{"id":50,"slug":51,"title":52,"cover_image":53,"image_url":53,"created_at":54,"category":13},"b400fb5d-3c21-4a6a-8383-988225159548","sparse-autoencoders-set-level-instability-en","Sparse Autoencoders Don’t Behave Like Feature Bags","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1786518181029-jyb4.png","2026-08-12T07:02:32.315345+00:00",{"id":56,"slug":57,"title":58,"cover_image":59,"image_url":59,"created_at":60,"category":13},"605dd415-e62d-455a-bb4b-d1d2fa487c1b","convawg-controlled-vawg-dialogue-generation-en","ConVAWG generates controlled VAWG dialogues","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1786516377348-lc87.png","2026-08-12T06:32:30.834532+00:00",{"id":62,"slug":63,"title":64,"cover_image":65,"image_url":65,"created_at":66,"category":13},"4c94994e-d58b-4f24-a480-ad026fd60e04","surgical-wam-video-pretraining-robot-control-en","Surgical WAM uses video to train robot control","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1786514579882-zhmr.png","2026-08-12T06:02:32.763172+00:00",{"id":68,"slug":69,"title":70,"cover_image":71,"image_url":71,"created_at":72,"category":13},"6d197f27-628f-4a63-883d-81a0d9f5c4b5","swe-bench-verified-model-leaderboard-limit-en","SWE-bench Verified has stopped being a clean model leaderboard","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1786498365866-w6ew.png","2026-08-12T01:32:20.055621+00:00",[74,79,84,89,94,99,104,109,114,119],{"id":75,"slug":76,"title":77,"created_at":78},"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":80,"slug":81,"title":82,"created_at":83},"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":85,"slug":86,"title":87,"created_at":88},"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":90,"slug":91,"title":92,"created_at":93},"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":95,"slug":96,"title":97,"created_at":98},"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":100,"slug":101,"title":102,"created_at":103},"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":105,"slug":106,"title":107,"created_at":108},"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":110,"slug":111,"title":112,"created_at":113},"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":115,"slug":116,"title":117,"created_at":118},"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":120,"slug":121,"title":122,"created_at":123},"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"]