{"$schema": "https://iptc.org/std/ninjs/ninjs-schema_3.2.json", "uri": "urn:newsml:wireva.mib.news:item:1a5be9a7-322b-4e08-81bf-826a3b55cd28", "type": "text", "profile": "standard", "version": "1", "versioncreated": "2026-10-12T19:05:00.766968Z", "firstcreated": "2026-10-12T19:00:43.092868Z", "pubstatus": "usable", "urgency": 5, "language": "en", "headline": "Sakana AI's Self-Improving Model Boosts Task Performance but Cannot Detect Its Own Blind Spots", "byline": "Jana Hartmann", "located": "Germany", "copyrightholder": "Haydamax OÜ", "copyrightnotice": "Wireva / Hochland", "usageterms": "Monitoring item. Headline, short extract and the source link may be shown. Do not republish the full text.", "description_text": "Sakana AI's MASS method allows a language model to improve itself by solving, judging, and refining tasks, reaching a nearly 70 percent win rate against its base version after two rounds. However, when the model consistently misses the same errors across all roles, it can mistake them for progress and reinforce them during training.", "altids": {"mosaic": "1a5be9a7-322b-4e08-81bf-826a3b55cd28", "story": "0f87c716-257c-44a3-a66d-3a4174a37d9a", "slug": "sakana-ai-s-self-improving-model-boosts-task-performance-but-cannot-detect-its-own-blind-spots"}, "subject": [{"name": "Technology", "rel": "direct", "literal": "technology", "code": "medtop:13000000", "scheme": "http://cv.iptc.org/newscodes/mediatopic/", "uri": "http://cv.iptc.org/newscodes/mediatopic/13000000"}], "place": [{"name": "Germany", "literal": "de", "code": "DE", "scheme": "http://cv.iptc.org/newscodes/countrycodes/"}], "person": [], "organisation": [], "genre": [{"name": "Standard story", "literal": "standard"}], "service": [{"name": "Wireva", "literal": "wireva"}], "body_text": "Sakana AI's MASS method allows a language model to improve itself by solving, judging, and refining tasks, reaching a nearly 70 percent win rate against its base version after two rounds. However, when the model consistently misses the same errors across all roles, it can mistake", "extra_body_note": "Extract only — see rights.class = link_only.", "associations": {"featuremedia": {"uri": "urn:newsml:wireva.mib.news:media:ff518a1a-c826-424e-840c-9ba92e80de59", "type": "picture", "version": "1", "pubstatus": "usable", "headline": "Sakana AI's self-improving model gets better at solving and judging, but can't catch its own blind spots", "description_text": "Sakana AI's Self-Improving Model Boosts Task Performance but Cannot Detect Its Own Blind Spots", "copyrightholder": "the-decoder.com", "copyrightnotice": "Sakana AI's self-improving model gets better at solving and judging, but can't catch its own blind spots", "usageterms": "Picture belongs to the source publication: address and terms are delivered, the file is not.", "altids": {"wireva": "ff518a1a-c826-424e-840c-9ba92e80de59"}, "extra": {"wireva:delivery_mode": "link_only", "wireva:rights": {"licence": "", "licence_url": "", "licence_raw": "RSS source", "licence_verified_at": "2026-10-12T19:05:00.786344Z", "attribution_required": true, "attribution_text": "Sakana AI's self-improving model gets better at solving and judging, but can't catch its own blind spots", "redistribution_allowed": false, "commercial_use_allowed": false, "modifications_allowed": false, "share_alike_required": false}, "wireva:provenance": {"creator": "", "source_name": "the-decoder.com", "source_url": "https://the-decoder.com/wp-content/uploads/2026/10/mass-recursive-self-improvement-optimization-nano-banana-pro.jpg", "source_page_url": "https://the-decoder.com/wp-content/uploads/2026/10/mass-recursive-self-improvement-optimization-nano-banana-pro.jpg", "our_page_url": "https://hochland.uno/technology/sakana-ai-s-self-improving-model-boosts-task-performance-but-cannot-detect-its-own-blind-spots/"}, "wireva:role": "lead"}, "renditions": {"source": {"href": "https://the-decoder.com/wp-content/uploads/2026/10/mass-recursive-self-improvement-optimization-nano-banana-pro.jpg", "mimetype": "image/jpeg"}}}}, "extra": {"wireva:story": {"uri": "urn:newsml:wireva.mib.news:story:0f87c716-257c-44a3-a66d-3a4174a37d9a", "slug": "sakana-ais-self-improving-model-boosts-task-performance-but-cannot-detect-its-own-blind-spots-20261012", "status": "developing", "languages": ["en"], "item_count": 1, "is_breaking": false}, "wireva:rights": {"class": "link_only", "redistribution_allowed": false, "commercial_use_allowed": false, "modifications_allowed": true, "share_alike_required": false, "attribution_required": true, "attribution_text": "Wireva / Hochland", "canonical_required": true, "canonical_url": "https://hochland.uno/technology/sakana-ai-s-self-improving-model-boosts-task-performance-but-cannot-detect-its-own-blind-spots/", "linkback_required": true, "licence": "", "licence_url": "", "licensor": "Haydamax OÜ", "valid_until": null, "excerpt_max_chars": 280, "decided_by_rule": "derivation_failed", "needs_review": true}, "wireva:provenance": {"production_method": "automated", "ai_stages": {"editing": true, "drafting": true, "headline": true, "discovery": true, "classification": true, "image_selection": true, "source_collection": true, "cross_source_comparison": true}, "model": "", "human_reviewed": false, "reviewed_by": "", "reviewed_at": null, "editorial_control": true, "editorial_responsibility": "Haydamax OÜ (Wireva)", "factchecked": false, "verification_status": "single_source", "independent_confirmations": 0, "official_source_used": false, "used_press_release": false, "public_interest": false, "ai_disclosure_required": false, "ai_disclosure_text": "", "independence_check": {"status": "failed", "score": 0.833, "checked_at": "2026-10-12T19:05:00.782652Z"}}, "wireva:publication": {"name": "Hochland", "domain": "hochland.uno", "country": "DE", "url": "https://hochland.uno/technology/sakana-ai-s-self-improving-model-boosts-task-performance-but-cannot-detect-its-own-blind-spots/"}, "wireva:links": {"self": "https://wireva.mib.news/api/v1/items/1a5be9a7-322b-4e08-81bf-826a3b55cd28", "html": "https://wireva.mib.news/item/1a5be9a7-322b-4e08-81bf-826a3b55cd28/sakana-ai-s-self-improving-model-boosts-task-performance-but-cannot-detect-its-own-blind-spots/", "story": "https://wireva.mib.news/story/sakana-ais-self-improving-model-boosts-task-performance-but-cannot-detect-its-own-blind-spots-20261012/"}, "wireva:metrics": {"word_count": 434, "char_count": 2871, "reading_time_seconds": 205}}}