{"id":"b434e4013d9220f4a526d55e0aedc92d27e22de9b7505e7937f1f6e821a7870a","title":"Long-Context Generation Is a Sampling Problem (When You Turn Up the Heat)","argument":"The authors argue that, at high temperature, the sampler—the rule that chooses the next token from a model’s probabilities—can determine whether long generation holds up, rather than more training being the only answer. They tested 29 open-weight models on ten benchmarks and needle retrieval out to one million tokens, including coding agents, long-form writing and long-context reading. In their results, samplers that keep the candidate-token pool small and steady as temperature rises retained accuracy; those admitting low-probability tokens lost it. At temperature 2, the authors report that coding agents using top-p, top-k or Top-H resolved none of 100 GitHub issues, while top-nsigma resolved 46 to 62 on three Qwen3.5 models. Across benchmarks, top-nsigma retained a median 100% of its temperature-1 score, compared with 26% for top-p. The authors link this difference to pool size: in one Llama-3.1-8B story, top-p’s pool grew from 4 tokens at temperature 1 to 51,845 at temperature 2, while top-nsigma stayed at 3. They also report that fixing the pool at temperature 1 stopped Top-H, min-p and p-less from collapsing when temperature was applied afterward. At temperature 1, most samplers performed about the same. The evidence presented here is limited to the tested models and tasks; whether the pattern extends beyond them remains open.","topicIds":["intelligence-and-discovery"],"domain":"science","claimType":"attributed reading","limits":"The abstract reports results from 29 open-weight models and the specified benchmarks and tasks; it does not establish whether the findings extend beyond those tests. The supplied evidence is the abstract and does not provide detailed evaluation methods or results for each benchmark.","speaker":{"name":"Roush, Allen G., Nguyen, Minh Nhat, Shwartz-Ziv, Ravid, Goldfeder, Judah, Basu, Sanjay","roleAtSource":"Article author"},"source":{"title":"Long-Context Generation Is a Sampling Problem (When You Turn Up the Heat)","publisher":"arXiv cs.AI","format":"Article","url":"https://arxiv.org/abs/2611.22148","publishedAt":"2026-11-17","dateBasis":"Date basis not provided."},"evidence":[{"text":"The authors argue that sampler choice at high temperature can determine whether long generation holds up.","excerpt":"At high temperature, The sampler, i.e. the rule that picks the next token out of the model's probabilities, decides whether long and agentic generation survives and/or thrives.","url":"https://arxiv.org/abs/2611.22148","relationship":"Source-stated"},{"text":"The authors tested 29 open-weight models across ten benchmarks and needle retrieval out to one million tokens.","excerpt":"We test 29 open-weight models from 1B to 122B parameters on ten established benchmarks plus needle retrieval out to one million tokens","url":"https://arxiv.org/abs/2611.22148","relationship":"Source-stated"},{"text":"At temperature 2, the authors report a large difference in coding-agent results between the named samplers on three Qwen3.5 models.","excerpt":"At temperature 2, coding agents running top-p, top-k or Top-H resolve 0 of 100 GitHub issues, while the same agents running top-nsigma resolve 46 to 62","url":"https://arxiv.org/abs/2611.22148","relationship":"Source-stated"},{"text":"The authors report that candidate-pool size predicts judged quality across the tested sampler settings and temperatures.","excerpt":"Across 133 combinations of sampler setting and temperature, the typical pool size predicts judged quality (Spearman correlation of -0.77)","url":"https://arxiv.org/abs/2611.22148","relationship":"Source-stated"},{"text":"At temperature 1, most samplers performed about the same, according to the authors.","excerpt":"At the default temperature of 1 most samplers perform about the same","url":"https://arxiv.org/abs/2611.22148","relationship":"Source-stated"}],"sourceMedia":{"status":"not_found","sourceUrl":"https://arxiv.org/abs/2611.22148","checkedAt":"2026-11-18T15:32:09.710Z","items":[],"note":"No suitable media was found by the supported page extraction."},"assumptions":[],"openQuestion":{"attribution":"Open question","text":"Do these results hold for open-weight models and generation tasks beyond the ones tested?"},"whatWouldChangeView":{"attribution":"Change condition not recorded","text":""},"editorialComparisons":[],"live":{"title":"Long-Context Generation Is a Sampling Problem (When You Turn Up the Heat)","summary":"The authors argue that, at high temperature, the sampler—the rule that chooses the next token from a model’s probabilities—can determine whether long generation holds up, rather than more training being the only answer. They tested 29 open-weight models on ten benchmarks and needle retrieval out to one million tokens, including coding agents, long-form writing and long-context reading. In their results, samplers that keep the candidate-token pool small and steady as temperature rises retained accuracy; those admitting low-probability tokens lost it. At temperature 2, the authors report that coding agents using top-p, top-k or Top-H resolved none of 100 GitHub issues, while top-nsigma resolved 46 to 62 on three Qwen3.5 models. Across benchmarks, top-nsigma retained a median 100% of its temperature-1 score, compared with 26% for top-p. The authors link this difference to pool size: in one Llama-3.1-8B story, top-p’s pool grew from 4 tokens at temperature 1 to 51,845 at temperature 2, while top-nsigma stayed at 3. They also report that fixing the pool at temperature 1 stopped Top-H, min-p and p-less from collapsing when temperature was applied afterward. At temperature 1, most samplers performed about the same. The evidence presented here is limited to the tested models and tasks; whether the pattern extends beyond them remains open.","question":"Do these results hold for open-weight models and generation tasks beyond the ones tested?","claims":[{"text":"The authors argue that sampler choice at high temperature can determine whether long generation holds up.","excerpt":"At high temperature, The sampler, i.e. the rule that picks the next token out of the model's probabilities, decides whether long and agentic generation survives and/or thrives.","support":"source-stated"},{"text":"The authors tested 29 open-weight models across ten benchmarks and needle retrieval out to one million tokens.","excerpt":"We test 29 open-weight models from 1B to 122B parameters on ten established benchmarks plus needle retrieval out to one million tokens","support":"source-stated"},{"text":"At temperature 2, the authors report a large difference in coding-agent results between the named samplers on three Qwen3.5 models.","excerpt":"At temperature 2, coding agents running top-p, top-k or Top-H resolve 0 of 100 GitHub issues, while the same agents running top-nsigma resolve 46 to 62","support":"source-stated"},{"text":"The authors report that candidate-pool size predicts judged quality across the tested sampler settings and temperatures.","excerpt":"Across 133 combinations of sampler setting and temperature, the typical pool size predicts judged quality (Spearman correlation of -0.77)","support":"source-stated"},{"text":"At temperature 1, most samplers performed about the same, according to the authors.","excerpt":"At the default temperature of 1 most samplers perform about the same","support":"source-stated"}],"limitations":["The abstract reports results from 29 open-weight models and the specified benchmarks and tasks; it does not establish whether the findings extend beyond those tests.","The supplied evidence is the abstract and does not provide detailed evaluation methods or results for each benchmark."],"mediaId":"a2d955f0dbea87d3da5a6f0eb91b63dc1cac50403c8742919d772860533b207c","correctionReason":null,"id":"b434e4013d9220f4a526d55e0aedc92d27e22de9b7505e7937f1f6e821a7870a","publicationId":"2b216d3a-39c2-4db4-8a19-52343885f679","version":1,"sourceUrl":"https://arxiv.org/abs/2611.22148","sourceTitle":"Long-Context Generation Is a Sampling Problem (When You Turn Up the Heat)","publisher":"arXiv cs.AI","publishedAt":null,"originalPublished":{"value":"2026-11-17","precision":"day","basis":"publisher_citation_date"},"firstObservedAt":"2026-11-18T05:13:57.783Z","createdAt":"2026-11-18T15:01:34.080Z","historical":false,"entityIds":["arxiv"],"topicIds":["intelligence-and-discovery"],"media":[{"id":"a2d955f0dbea87d3da5a6f0eb91b63dc1cac50403c8742919d772860533b207c","kind":"article","url":"https://arxiv.org/abs/2611.22148","sourceUrl":"https://arxiv.org/abs/2611.22148","label":"Read original","provenance":"source_link","verification":{"status":"verified","checkedAt":"2026-11-18T14:49:27.620Z","method":"acquired_source","contentType":"text/html"},"recommendationReason":"Recommended by Jev for understanding this source; checked by Luna."}],"sourceMedia":{"status":"not_found","sourceUrl":"https://arxiv.org/abs/2611.22148","checkedAt":"2026-11-18T15:32:09.710Z","items":[],"note":"No suitable media was found by the supported page extraction."},"editorialDecision":{"outcome":"accept","reasonCodes":["material_development"],"rubricVersion":"stream-admission-v9","mediaId":"a2d955f0dbea87d3da5a6f0eb91b63dc1cac50403c8742919d772860533b207c","confidence":0.56,"provider":"typesafe","model":"jev-1.13.0","inputHash":"b30987a126e3306b09596771d74e4ec081d4947171064ccd727f81dd93d11f76","decidedAt":"2026-11-18T14:49:29.358Z"},"provenance":{"textHash":"746a140ba63c658c2115097dbfa87d8a74e774219ec84d18870a8add17c43bbf","observationId":"94601fd6313cc467c10d028e746043694038b8ac6c6be2d042af996964da5a90","retrievedAt":"2026-11-18T14:49:27.375Z","kind":"publisher-page","preparer":{"model":"gpt-6-luna","runId":"458336b9-a2d5-4635-806c-b191aa22d33f"},"reviewer":{"model":"gpt-6-luna","runId":"966f7b5a-6714-4e5f-802c-8feaf87fd9ea"}},"audit":{"status":"uncertain","reviewedAt":"2026-11-18T15:01:34.080Z","model":"anthropic/claude-opus-5.5"},"sourceAttribution":{"authors":[{"name":"Roush, Allen G.","url":null,"kind":"unknown"},{"name":"Nguyen, Minh Nhat","url":null,"kind":"unknown"},{"name":"Shwartz-Ziv, Ravid","url":null,"kind":"unknown"},{"name":"Goldfeder, Judah","url":null,"kind":"unknown"},{"name":"Basu, Sanjay","url":null,"kind":"unknown"}],"authorRole":"article-author","hosts":[],"status":"stated","authorRoster":{"truncated":false,"omitted":0,"assertionGroupsOmitted":0},"participation":[{"actor":{"name":"Roush, Allen G.","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Roush, Allen G.","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2611.22148","locator":"head/meta[name=author|citation_author]"},"assertionId":"7032f5f521b9b2aa2ce11dc6fc44b2197a56e40b2a36c8d7f69a17cba00b07c9:participation:0","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-18T15:32:09.710Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"7032f5f521b9b2aa2ce11dc6fc44b2197a56e40b2a36c8d7f69a17cba00b07c9","sourceUrl":"https://arxiv.org/abs/2611.22148","contentSha256":"3a3e8fa11cd03774c972e0af062ffb428e0412421c2719d36dfd12983bbc5616","textHash":"746a140ba63c658c2115097dbfa87d8a74e774219ec84d18870a8add17c43bbf","observedAt":"2026-11-18T15:32:09.710Z"}},{"actor":{"name":"Nguyen, Minh Nhat","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Nguyen, Minh Nhat","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2611.22148","locator":"head/meta[name=author|citation_author]"},"assertionId":"7032f5f521b9b2aa2ce11dc6fc44b2197a56e40b2a36c8d7f69a17cba00b07c9:participation:1","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-18T15:32:09.710Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"7032f5f521b9b2aa2ce11dc6fc44b2197a56e40b2a36c8d7f69a17cba00b07c9","sourceUrl":"https://arxiv.org/abs/2611.22148","contentSha256":"3a3e8fa11cd03774c972e0af062ffb428e0412421c2719d36dfd12983bbc5616","textHash":"746a140ba63c658c2115097dbfa87d8a74e774219ec84d18870a8add17c43bbf","observedAt":"2026-11-18T15:32:09.710Z"}},{"actor":{"name":"Shwartz-Ziv, Ravid","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Shwartz-Ziv, Ravid","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2611.22148","locator":"head/meta[name=author|citation_author]"},"assertionId":"7032f5f521b9b2aa2ce11dc6fc44b2197a56e40b2a36c8d7f69a17cba00b07c9:participation:2","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-18T15:32:09.710Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"7032f5f521b9b2aa2ce11dc6fc44b2197a56e40b2a36c8d7f69a17cba00b07c9","sourceUrl":"https://arxiv.org/abs/2611.22148","contentSha256":"3a3e8fa11cd03774c972e0af062ffb428e0412421c2719d36dfd12983bbc5616","textHash":"746a140ba63c658c2115097dbfa87d8a74e774219ec84d18870a8add17c43bbf","observedAt":"2026-11-18T15:32:09.710Z"}},{"actor":{"name":"Goldfeder, Judah","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Goldfeder, Judah","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2611.22148","locator":"head/meta[name=author|citation_author]"},"assertionId":"7032f5f521b9b2aa2ce11dc6fc44b2197a56e40b2a36c8d7f69a17cba00b07c9:participation:3","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-18T15:32:09.710Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"7032f5f521b9b2aa2ce11dc6fc44b2197a56e40b2a36c8d7f69a17cba00b07c9","sourceUrl":"https://arxiv.org/abs/2611.22148","contentSha256":"3a3e8fa11cd03774c972e0af062ffb428e0412421c2719d36dfd12983bbc5616","textHash":"746a140ba63c658c2115097dbfa87d8a74e774219ec84d18870a8add17c43bbf","observedAt":"2026-11-18T15:32:09.710Z"}},{"actor":{"name":"Basu, Sanjay","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Basu, Sanjay","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2611.22148","locator":"head/meta[name=author|citation_author]"},"assertionId":"7032f5f521b9b2aa2ce11dc6fc44b2197a56e40b2a36c8d7f69a17cba00b07c9:participation:4","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-18T15:32:09.710Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"7032f5f521b9b2aa2ce11dc6fc44b2197a56e40b2a36c8d7f69a17cba00b07c9","sourceUrl":"https://arxiv.org/abs/2611.22148","contentSha256":"3a3e8fa11cd03774c972e0af062ffb428e0412421c2719d36dfd12983bbc5616","textHash":"746a140ba63c658c2115097dbfa87d8a74e774219ec84d18870a8add17c43bbf","observedAt":"2026-11-18T15:32:09.710Z"}}],"participationCoverage":{"total":5,"returned":5,"complete":true,"fullAssertions":"source-metadata-endpoint"},"distribution":null,"publisher":null,"sourceUrl":"https://arxiv.org/abs/2611.22148","snapshotId":"7032f5f521b9b2aa2ce11dc6fc44b2197a56e40b2a36c8d7f69a17cba00b07c9","contentSha256":"3a3e8fa11cd03774c972e0af062ffb428e0412421c2719d36dfd12983bbc5616","textHash":"746a140ba63c658c2115097dbfa87d8a74e774219ec84d18870a8add17c43bbf","observedAt":"2026-11-18T15:32:09.710Z","basis":"publisher-page-metadata","sourceLinks":44,"linksTruncated":false,"metadataUrl":"/api/pipeline/stream/b434e4013d9220f4a526d55e0aedc92d27e22de9b7505e7937f1f6e821a7870a/source"}},"canonicalUrl":"https://escapemeridian.com/readings/long-context-generation-is-a-sampling-problem-when-you-turn-up-b434e401","formats":{"html":"/readings/long-context-generation-is-a-sampling-problem-when-you-turn-up-b434e401","markdown":"/readings/long-context-generation-is-a-sampling-problem-when-you-turn-up-b434e401.md","json":"/readings/long-context-generation-is-a-sampling-problem-when-you-turn-up-b434e401.json"},"attribution":"Headline and summary by Escape Meridian, prepared by Luna from the attributed source. A source check is not independent confirmation of the source’s claims.","maintainedGuideLinks":[],"conditions":[],"readerContext":{"version":"reader-context-v1","attribution":"Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.","terms":[{"id":"term-open-weights","name":"Open weights","explanation":"A model whose trained numbers are published so others can download, run and inspect it; the license and training data may still be restricted.","referenceUrl":"https://opensource.org/ai/open-weights"},{"id":"term-benchmark","name":"Benchmark","explanation":"A defined set of tests used to compare systems. Results depend on the tasks and measures selected.","referenceUrl":"https://arxiv.org/abs/2211.09110"},{"id":"term-tokens","name":"Model tokens","explanation":"Small chunks of text, often parts of words, that a language model reads and writes; usage and prices are often counted in tokens.","referenceUrl":"https://developers.google.com/machine-learning/glossary#token"}],"entities":[{"id":"arxiv","name":"arXiv","kind":"organization","context":"arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.","referenceUrls":["https://arxiv.org/","https://info.arxiv.org/help/api/index.html"],"path":"/entities/arxiv","json":"/entities/arxiv.json","markdown":"/entities/arxiv.md"}]},"researchQuestions":[],"withdrawn":false,"references":{"schemaVersion":1,"target":{"type":"reading","id":"b434e4013d9220f4a526d55e0aedc92d27e22de9b7505e7937f1f6e821a7870a","revision":"2b216d3a-39c2-4db4-8a19-52343885f679","revisionKind":"publication-id","title":"Long-Context Generation Is a Sampling Problem (When You Turn Up the Heat)","url":"https://escapemeridian.com/readings/long-context-generation-is-a-sampling-problem-when-you-turn-up-b434e401"},"source":{"type":"source","id":"https://arxiv.org/abs/2611.22148","revision":"746a140ba63c658c2115097dbfa87d8a74e774219ec84d18870a8add17c43bbf","revisionKind":"sha256-source-text","title":"Long-Context Generation Is a Sampling Problem (When You Turn Up the Heat)","url":"https://arxiv.org/abs/2611.22148"},"status":"available","items":[],"next":null,"scope":"Explicit citations in public readings to this reading. Shared topics are not citations. Other record types need an explicit reference adapter."},"revisions":[{"id":"b434e4013d9220f4a526d55e0aedc92d27e22de9b7505e7937f1f6e821a7870a","title":"Long-Context Generation Is a Sampling Problem (When You Turn Up the Heat)","argument":"The authors argue that, at high temperature, the sampler—the rule that chooses the next token from a model’s probabilities—can determine whether long generation holds up, rather than more training being the only answer. They tested 29 open-weight models on ten benchmarks and needle retrieval out to one million tokens, including coding agents, long-form writing and long-context reading. In their results, samplers that keep the candidate-token pool small and steady as temperature rises retained accuracy; those admitting low-probability tokens lost it. At temperature 2, the authors report that coding agents using top-p, top-k or Top-H resolved none of 100 GitHub issues, while top-nsigma resolved 46 to 62 on three Qwen3.5 models. Across benchmarks, top-nsigma retained a median 100% of its temperature-1 score, compared with 26% for top-p. The authors link this difference to pool size: in one Llama-3.1-8B story, top-p’s pool grew from 4 tokens at temperature 1 to 51,845 at temperature 2, while top-nsigma stayed at 3. They also report that fixing the pool at temperature 1 stopped Top-H, min-p and p-less from collapsing when temperature was applied afterward. At temperature 1, most samplers performed about the same. The evidence presented here is limited to the tested models and tasks; whether the pattern extends beyond them remains open.","topicIds":["intelligence-and-discovery"],"domain":"science","claimType":"attributed reading","limits":"The abstract reports results from 29 open-weight models and the specified benchmarks and tasks; it does not establish whether the findings extend beyond those tests. The supplied evidence is the abstract and does not provide detailed evaluation methods or results for each benchmark.","speaker":{"name":"Roush, Allen G., Nguyen, Minh Nhat, Shwartz-Ziv, Ravid, Goldfeder, Judah, Basu, Sanjay","roleAtSource":"Article author"},"source":{"title":"Long-Context Generation Is a Sampling Problem (When You Turn Up the Heat)","publisher":"arXiv cs.AI","format":"Article","url":"https://arxiv.org/abs/2611.22148","publishedAt":"2026-11-17","dateBasis":"Date basis not provided."},"evidence":[{"text":"The authors argue that sampler choice at high temperature can determine whether long generation holds up.","excerpt":"At high temperature, The sampler, i.e. the rule that picks the next token out of the model's probabilities, decides whether long and agentic generation survives and/or thrives.","url":"https://arxiv.org/abs/2611.22148","relationship":"Source-stated"},{"text":"The authors tested 29 open-weight models across ten benchmarks and needle retrieval out to one million tokens.","excerpt":"We test 29 open-weight models from 1B to 122B parameters on ten established benchmarks plus needle retrieval out to one million tokens","url":"https://arxiv.org/abs/2611.22148","relationship":"Source-stated"},{"text":"At temperature 2, the authors report a large difference in coding-agent results between the named samplers on three Qwen3.5 models.","excerpt":"At temperature 2, coding agents running top-p, top-k or Top-H resolve 0 of 100 GitHub issues, while the same agents running top-nsigma resolve 46 to 62","url":"https://arxiv.org/abs/2611.22148","relationship":"Source-stated"},{"text":"The authors report that candidate-pool size predicts judged quality across the tested sampler settings and temperatures.","excerpt":"Across 133 combinations of sampler setting and temperature, the typical pool size predicts judged quality (Spearman correlation of -0.77)","url":"https://arxiv.org/abs/2611.22148","relationship":"Source-stated"},{"text":"At temperature 1, most samplers performed about the same, according to the authors.","excerpt":"At the default temperature of 1 most samplers perform about the same","url":"https://arxiv.org/abs/2611.22148","relationship":"Source-stated"}],"sourceMedia":{"status":"not_found","sourceUrl":"https://arxiv.org/abs/2611.22148","checkedAt":"2026-11-18T15:32:09.710Z","items":[],"note":"No suitable media was found by the supported page extraction."},"assumptions":[],"openQuestion":{"attribution":"Open question","text":"Do these results hold for open-weight models and generation tasks beyond the ones tested?"},"whatWouldChangeView":{"attribution":"Change condition not recorded","text":""},"editorialComparisons":[],"live":{"title":"Long-Context Generation Is a Sampling Problem (When You Turn Up the Heat)","summary":"The authors argue that, at high temperature, the sampler—the rule that chooses the next token from a model’s probabilities—can determine whether long generation holds up, rather than more training being the only answer. They tested 29 open-weight models on ten benchmarks and needle retrieval out to one million tokens, including coding agents, long-form writing and long-context reading. In their results, samplers that keep the candidate-token pool small and steady as temperature rises retained accuracy; those admitting low-probability tokens lost it. At temperature 2, the authors report that coding agents using top-p, top-k or Top-H resolved none of 100 GitHub issues, while top-nsigma resolved 46 to 62 on three Qwen3.5 models. Across benchmarks, top-nsigma retained a median 100% of its temperature-1 score, compared with 26% for top-p. The authors link this difference to pool size: in one Llama-3.1-8B story, top-p’s pool grew from 4 tokens at temperature 1 to 51,845 at temperature 2, while top-nsigma stayed at 3. They also report that fixing the pool at temperature 1 stopped Top-H, min-p and p-less from collapsing when temperature was applied afterward. At temperature 1, most samplers performed about the same. The evidence presented here is limited to the tested models and tasks; whether the pattern extends beyond them remains open.","question":"Do these results hold for open-weight models and generation tasks beyond the ones tested?","claims":[{"text":"The authors argue that sampler choice at high temperature can determine whether long generation holds up.","excerpt":"At high temperature, The sampler, i.e. the rule that picks the next token out of the model's probabilities, decides whether long and agentic generation survives and/or thrives.","support":"source-stated"},{"text":"The authors tested 29 open-weight models across ten benchmarks and needle retrieval out to one million tokens.","excerpt":"We test 29 open-weight models from 1B to 122B parameters on ten established benchmarks plus needle retrieval out to one million tokens","support":"source-stated"},{"text":"At temperature 2, the authors report a large difference in coding-agent results between the named samplers on three Qwen3.5 models.","excerpt":"At temperature 2, coding agents running top-p, top-k or Top-H resolve 0 of 100 GitHub issues, while the same agents running top-nsigma resolve 46 to 62","support":"source-stated"},{"text":"The authors report that candidate-pool size predicts judged quality across the tested sampler settings and temperatures.","excerpt":"Across 133 combinations of sampler setting and temperature, the typical pool size predicts judged quality (Spearman correlation of -0.77)","support":"source-stated"},{"text":"At temperature 1, most samplers performed about the same, according to the authors.","excerpt":"At the default temperature of 1 most samplers perform about the same","support":"source-stated"}],"limitations":["The abstract reports results from 29 open-weight models and the specified benchmarks and tasks; it does not establish whether the findings extend beyond those tests.","The supplied evidence is the abstract and does not provide detailed evaluation methods or results for each benchmark."],"mediaId":"a2d955f0dbea87d3da5a6f0eb91b63dc1cac50403c8742919d772860533b207c","correctionReason":null,"id":"b434e4013d9220f4a526d55e0aedc92d27e22de9b7505e7937f1f6e821a7870a","publicationId":"2b216d3a-39c2-4db4-8a19-52343885f679","version":1,"sourceUrl":"https://arxiv.org/abs/2611.22148","sourceTitle":"Long-Context Generation Is a Sampling Problem (When You Turn Up the Heat)","publisher":"arXiv cs.AI","publishedAt":null,"originalPublished":{"value":"2026-11-17","precision":"day","basis":"publisher_citation_date"},"firstObservedAt":"2026-11-18T05:13:57.783Z","createdAt":"2026-11-18T15:01:34.080Z","historical":false,"entityIds":["arxiv"],"topicIds":["intelligence-and-discovery"],"media":[{"id":"a2d955f0dbea87d3da5a6f0eb91b63dc1cac50403c8742919d772860533b207c","kind":"article","url":"https://arxiv.org/abs/2611.22148","sourceUrl":"https://arxiv.org/abs/2611.22148","label":"Read original","provenance":"source_link","verification":{"status":"verified","checkedAt":"2026-11-18T14:49:27.620Z","method":"acquired_source","contentType":"text/html"},"recommendationReason":"Recommended by Jev for understanding this source; checked by Luna."}],"sourceMedia":{"status":"not_found","sourceUrl":"https://arxiv.org/abs/2611.22148","checkedAt":"2026-11-18T15:32:09.710Z","items":[],"note":"No suitable media was found by the supported page extraction."},"editorialDecision":{"outcome":"accept","reasonCodes":["material_development"],"rubricVersion":"stream-admission-v9","mediaId":"a2d955f0dbea87d3da5a6f0eb91b63dc1cac50403c8742919d772860533b207c","confidence":0.56,"provider":"typesafe","model":"jev-1.13.0","inputHash":"b30987a126e3306b09596771d74e4ec081d4947171064ccd727f81dd93d11f76","decidedAt":"2026-11-18T14:49:29.358Z"},"provenance":{"textHash":"746a140ba63c658c2115097dbfa87d8a74e774219ec84d18870a8add17c43bbf","observationId":"94601fd6313cc467c10d028e746043694038b8ac6c6be2d042af996964da5a90","retrievedAt":"2026-11-18T14:49:27.375Z","kind":"publisher-page","preparer":{"model":"gpt-6-luna","runId":"458336b9-a2d5-4635-806c-b191aa22d33f"},"reviewer":{"model":"gpt-6-luna","runId":"966f7b5a-6714-4e5f-802c-8feaf87fd9ea"}},"audit":{"status":"uncertain","reviewedAt":"2026-11-18T15:01:34.080Z","model":"anthropic/claude-opus-5.5"},"sourceAttribution":{"authors":[{"name":"Roush, Allen G.","url":null,"kind":"unknown"},{"name":"Nguyen, Minh Nhat","url":null,"kind":"unknown"},{"name":"Shwartz-Ziv, Ravid","url":null,"kind":"unknown"},{"name":"Goldfeder, Judah","url":null,"kind":"unknown"},{"name":"Basu, Sanjay","url":null,"kind":"unknown"}],"authorRole":"article-author","hosts":[],"status":"stated","authorRoster":{"truncated":false,"omitted":0,"assertionGroupsOmitted":0},"participation":[{"actor":{"name":"Roush, Allen G.","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Roush, Allen G.","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2611.22148","locator":"head/meta[name=author|citation_author]"},"assertionId":"7032f5f521b9b2aa2ce11dc6fc44b2197a56e40b2a36c8d7f69a17cba00b07c9:participation:0","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-18T15:32:09.710Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"7032f5f521b9b2aa2ce11dc6fc44b2197a56e40b2a36c8d7f69a17cba00b07c9","sourceUrl":"https://arxiv.org/abs/2611.22148","contentSha256":"3a3e8fa11cd03774c972e0af062ffb428e0412421c2719d36dfd12983bbc5616","textHash":"746a140ba63c658c2115097dbfa87d8a74e774219ec84d18870a8add17c43bbf","observedAt":"2026-11-18T15:32:09.710Z"}},{"actor":{"name":"Nguyen, Minh Nhat","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Nguyen, Minh Nhat","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2611.22148","locator":"head/meta[name=author|citation_author]"},"assertionId":"7032f5f521b9b2aa2ce11dc6fc44b2197a56e40b2a36c8d7f69a17cba00b07c9:participation:1","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-18T15:32:09.710Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"7032f5f521b9b2aa2ce11dc6fc44b2197a56e40b2a36c8d7f69a17cba00b07c9","sourceUrl":"https://arxiv.org/abs/2611.22148","contentSha256":"3a3e8fa11cd03774c972e0af062ffb428e0412421c2719d36dfd12983bbc5616","textHash":"746a140ba63c658c2115097dbfa87d8a74e774219ec84d18870a8add17c43bbf","observedAt":"2026-11-18T15:32:09.710Z"}},{"actor":{"name":"Shwartz-Ziv, Ravid","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Shwartz-Ziv, Ravid","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2611.22148","locator":"head/meta[name=author|citation_author]"},"assertionId":"7032f5f521b9b2aa2ce11dc6fc44b2197a56e40b2a36c8d7f69a17cba00b07c9:participation:2","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-18T15:32:09.710Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"7032f5f521b9b2aa2ce11dc6fc44b2197a56e40b2a36c8d7f69a17cba00b07c9","sourceUrl":"https://arxiv.org/abs/2611.22148","contentSha256":"3a3e8fa11cd03774c972e0af062ffb428e0412421c2719d36dfd12983bbc5616","textHash":"746a140ba63c658c2115097dbfa87d8a74e774219ec84d18870a8add17c43bbf","observedAt":"2026-11-18T15:32:09.710Z"}},{"actor":{"name":"Goldfeder, Judah","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Goldfeder, Judah","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2611.22148","locator":"head/meta[name=author|citation_author]"},"assertionId":"7032f5f521b9b2aa2ce11dc6fc44b2197a56e40b2a36c8d7f69a17cba00b07c9:participation:3","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-18T15:32:09.710Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"7032f5f521b9b2aa2ce11dc6fc44b2197a56e40b2a36c8d7f69a17cba00b07c9","sourceUrl":"https://arxiv.org/abs/2611.22148","contentSha256":"3a3e8fa11cd03774c972e0af062ffb428e0412421c2719d36dfd12983bbc5616","textHash":"746a140ba63c658c2115097dbfa87d8a74e774219ec84d18870a8add17c43bbf","observedAt":"2026-11-18T15:32:09.710Z"}},{"actor":{"name":"Basu, Sanjay","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Basu, Sanjay","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2611.22148","locator":"head/meta[name=author|citation_author]"},"assertionId":"7032f5f521b9b2aa2ce11dc6fc44b2197a56e40b2a36c8d7f69a17cba00b07c9:participation:4","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-18T15:32:09.710Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"7032f5f521b9b2aa2ce11dc6fc44b2197a56e40b2a36c8d7f69a17cba00b07c9","sourceUrl":"https://arxiv.org/abs/2611.22148","contentSha256":"3a3e8fa11cd03774c972e0af062ffb428e0412421c2719d36dfd12983bbc5616","textHash":"746a140ba63c658c2115097dbfa87d8a74e774219ec84d18870a8add17c43bbf","observedAt":"2026-11-18T15:32:09.710Z"}}],"participationCoverage":{"total":5,"returned":5,"complete":true,"fullAssertions":"source-metadata-endpoint"},"distribution":null,"publisher":null,"sourceUrl":"https://arxiv.org/abs/2611.22148","snapshotId":"7032f5f521b9b2aa2ce11dc6fc44b2197a56e40b2a36c8d7f69a17cba00b07c9","contentSha256":"3a3e8fa11cd03774c972e0af062ffb428e0412421c2719d36dfd12983bbc5616","textHash":"746a140ba63c658c2115097dbfa87d8a74e774219ec84d18870a8add17c43bbf","observedAt":"2026-11-18T15:32:09.710Z","basis":"publisher-page-metadata","sourceLinks":44,"linksTruncated":false,"metadataUrl":"/api/pipeline/stream/b434e4013d9220f4a526d55e0aedc92d27e22de9b7505e7937f1f6e821a7870a/source"}}}],"audits":[{"publicationId":"2b216d3a-39c2-4db4-8a19-52343885f679","verdict":"uncertain","findings":[{"dimension":"attribution","status":"pass","explanation":"The reading consistently attributes the argument and results to the paper's authors, and the arXiv source lists them. It does not present their findings as Escape Meridian's conclusions.","excerpt":"Authors:Allen G. Roush, Minh Nhat Nguyen, Ravid Shwartz-Ziv, Judah Goldfeder, Sanjay Basu"},{"dimension":"summary","status":"uncertain","explanation":"The reading drops the problem the paper starts from: open-weight models are widely observed to get worse the longer they generate. Without it, 'whether long generation holds up' lacks a clear stake. 'Temperature' is central to the paper and its title but is never explained, so a general reader cannot follow why raising it enlarges the candidate pool. Smallest repair: open with the degradation problem, and give a plain-language gloss or tooltip for temperature as the setting that controls randomness. A tooltip for top-p or top-nsigma would also help. 'Lost it' softens the source's 'fall apart'.","excerpt":"Open-weight language models are widely observed to get worse the longer they generate."},{"dimension":"date","status":"pass","explanation":"The displayed original date is 2026-11-17, which matches the v1 submission date. The feed date of 2026-11-18 is later and is correctly not used.","excerpt":"[Submitted on 17 Nov 2026]"},{"dimension":"support","status":"pass","explanation":"The figures in the reading match the abstract: 29 models, ten benchmarks, 0 of 100 versus 46 to 62 issues resolved, median 100% versus 26%, pool sizes of 4 to 51,845 versus 3, Spearman correlation of -0.77, and the fix of choosing the pool at temperature 1. The phrase 'can determine' is a cautious rendering of the source's 'decides'. The limitations and the closing question do not overreach beyond the evidence.","excerpt":"Over all benchmarks top-nsigma keeps a median 100% of its temperature-1 score where top-p keeps 26%."},{"dimension":"media","status":"pass","explanation":"The selected media is the verified arXiv abstract page itself, so it is the direct original source and the easiest route to the paper.","excerpt":null}],"createdAt":"2026-11-18T15:01:34.080Z","model":"anthropic/claude-opus-5.5"}],"correctionUrl":"https://escapemeridian.com/corrections/"}