{"id":1180547,"date":"2026-08-01T07:26:54","date_gmt":"2026-08-01T14:26:54","guid":{"rendered":"https:\/\/research.codeghost.online\/en-us\/research\/publication\/trace-turn-level-reward-assignment-via-credit-estimation-for-long-horizon-agents\/"},"modified":"2026-08-03T08:38:10","modified_gmt":"2026-08-03T15:38:10","slug":"trace-turn-level-reward-assignment-via-credit-estimation-for-long-horizon-agents","status":"publish","type":"msr-research-item","link":"https:\/\/research.codeghost.online\/en-us\/research\/publication\/trace-turn-level-reward-assignment-via-credit-estimation-for-long-horizon-agents\/","title":{"rendered":"TRACE: Turn-level Reward Assignment via Credit Estimation for Long-Horizon Agents"},"content":{"rendered":"\n\n\n<p class=\"wp-block-paragraph\">Multi-turn agents solve complex tasks through extended sequences of tool interactions before producing a final answer, making credit assignment a fundamental challenge during post-training. Outcome rewards provide reliable supervision for short-horizon reasoning, but become sparse and high-variance as trajectories grow to tens or hundreds of tool calls. They can also be misleading: a failed rollout may contain many useful actions that move the agent closer to the goal, yet outcome-only training assigns them the same negative advantage as the eventual mistake. We propose TRACE (Turn-level Reward Assignment via Credit Estimation), a dense credit-assignment method for agentic reinforcement learning. TRACE represents rollouts as state transitions at tool-call boundaries, obtains gold-answer log-probabilities from a frozen reference model, transforms them into log-ratio state values, and derives per-action rewards as Temporal-Difference changes in those values. This requires no additional critic or process-label training, and its one-step log-ratio TD component telescopes across redundant tool calls. On long-horizon complex search, TRACE substantially improves base-model tool-use ability using pure RL, without a cold-start supervised fine-tuning stage, an agentic mid-training stage, or training on live-web data. On the closed-web BrowseComp-Plus benchmark, it raises Qwen3-4B from <math><mn>7.2<\/mn><\/math> to <math><mn>35.6<\/mn><\/math> and Qwen3-30B-A3B from <math><mn>8.4<\/mn><\/math> to <math><mn>42.6<\/mn><\/math>. The learned search behavior also transfers to open-web benchmarks, and the learning curves show earlier improvement and faster convergence during RL training.<\/p>\n","protected":false},"excerpt":{"rendered":"<p>Multi-turn agents solve complex tasks through extended sequences of tool interactions before producing a final answer, making credit assignment a fundamental challenge during post-training. Outcome rewards provide reliable supervision for short-horizon reasoning, but become sparse and high-variance as trajectories grow to tens or hundreds of tool calls. They can also be misleading: a failed rollout [&hellip;]<\/p>\n","protected":false},"featured_media":0,"template":"","meta":{"msr-url-field":"","msr-podcast-episode":"","msrModifiedDate":"","msrModifiedDateEnabled":false,"ep_exclude_from_search":false,"_classifai_error":"","msr-author-ordering":[{"type":"text","value":"Leitian Tao","user_id":0},{"type":"user_nicename","value":"Baolin Peng","user_id":"43779"},{"type":"text","value":"Wenlin Yao","user_id":0},{"type":"user_nicename","value":"Tao Ge","user_id":"38850"},{"type":"user_nicename","value":"Hao Cheng","user_id":"39922"},{"type":"user_nicename","value":"Mike Hang Wang","user_id":"43951"},{"type":"user_nicename","value":"Jianfeng Gao","user_id":"32246"},{"type":"text","value":"Sharon Li","user_id":0}],"msr_publishername":"","msr_publisher_other":"","msr_booktitle":"","msr_chapter":"","msr_edition":"","msr_editors":"","msr_how_published":"arXiv","msr_isbn":"","msr_issue":"","msr_journal":"","msr_number":"","msr_organization":"","msr_pages_string":"","msr_page_range_start":"","msr_page_range_end":"","msr_series":"","msr_volume":"","msr_copyright":"","msr_conference_name":"","msr_doi":"","msr_arxiv_id":"2607.13988","msr_mag_id":"","msr_other_authors":"","msr_other_contributors":"","msr_speaker":"","msr_award":"","msr_affiliation":"","msr_institution":"","msr_host":"","msr_version":"","msr_duration":"","msr_release_tracker_id":"","msr_highlight_type":"","msr_date_display_format":"","msr_main_download_label":"","msr_external_link_label":"","msr_doi_label":"","msr_published_date":"2026-07-15","msr_startdate":"","msr_presentation_date":"","msr_highlight_text":"","msr_notes":"","msr_longbiography":"","msr_publicationurl":"","msr_external_url":"","msr_secondary_video_url":"","msr_conference_url":"","msr_journal_url":"","msr_year":2026,"msr_month":7,"msr_day":15,"msr_microsoftintellectualproperty":false,"msr_pub_id":"c9c2fb7719370179b4fac37154f53d8f12080cce","msr_publication_uploader":[{"type":"url","title":"https:\/\/arxiv.org\/abs\/2607.13988","label_id":243109,"id":false,"viewUrl":false}],"msr_related_uploader":[],"msr_original_fields_of_study":[],"msr_s2_paper_id":"c9c2fb7719370179b4fac37154f53d8f12080cce","msr_s2_pdf_url":"","msr_citation_count_updated":"","msr_citation_count":0,"msr_influential_citations":0,"msr_reference_count":89,"msr_s2_open_access":false,"msr_s2_author_ids":[],"msr_pub_ids":[{"provider":"s2","id":"c9c2fb7719370179b4fac37154f53d8f12080cce"},{"provider":"arxiv","id":"2607.13988"},{"provider":"corpusid","id":"290183464"}],"msr_hide_image_in_river":0,"footnotes":""},"msr-research-highlight":[],"research-area":[13556],"msr-publication-type":[270373],"msr-publisher":[],"msr-publication-cta":[],"msr-focus-area":[],"msr-locale":[268875],"msr-post-option":[],"msr-field-of-study":[246691,265497],"msr-conference":[],"msr-journal":[],"msr-impact-theme":[],"msr-pillar":[],"class_list":["post-1180547","msr-research-item","type-msr-research-item","status-publish","hentry","msr-research-area-artificial-intelligence","msr-locale-en_us","msr-field-of-study-computer-science","msr-field-of-study-machine-learning-296"],"msr_publishername":"","msr_edition":"","msr_affiliation":"","msr_published_date":"2026-07-15","msr_host":"","msr_duration":"","msr_version":"","msr_speaker":"","msr_other_contributors":"","msr_booktitle":"","msr_pages_string":"","msr_chapter":"","msr_isbn":"","msr_journal":"","msr_volume":"","msr_number":"","msr_editors":"","msr_series":"","msr_issue":"","msr_organization":"","msr_how_published":"arXiv","msr_notes":"","msr_highlight_text":"","msr_release_tracker_id":"","msr_original_fields_of_study":"","msr_download_urls":"","msr_external_url":"","msr_secondary_video_url":"","msr_longbiography":"","msr_microsoftintellectualproperty":0,"msr_main_download":"","msr_publicationurl":"","msr_doi":"","msr_publication_uploader":[{"type":"url","title":"https:\/\/arxiv.org\/abs\/2607.13988","label_id":243109,"id":false,"viewUrl":false}],"msr_related_uploader":[],"msr_citation_count":0,"msr_citation_count_updated":"","msr_s2_paper_id":"c9c2fb7719370179b4fac37154f53d8f12080cce","msr_influential_citations":0,"msr_reference_count":89,"msr_arxiv_id":"2607.13988","msr_s2_author_ids":[],"msr_s2_open_access":false,"msr_s2_pdf_url":null,"msr_attachments":[],"msr-author-ordering":[{"type":"text","value":"Leitian Tao","user_id":0,"rest_url":false},{"type":"user_nicename","value":"Baolin Peng","user_id":43779,"rest_url":"https:\/\/research.codeghost.online\/en-us\/research\/wp-json\/microsoft-research\/v1\/researchers?person=Baolin Peng"},{"type":"text","value":"Wenlin Yao","user_id":0,"rest_url":false},{"type":"user_nicename","value":"Tao Ge","user_id":38850,"rest_url":"https:\/\/research.codeghost.online\/en-us\/research\/wp-json\/microsoft-research\/v1\/researchers?person=Tao Ge"},{"type":"user_nicename","value":"Hao Cheng","user_id":39922,"rest_url":"https:\/\/research.codeghost.online\/en-us\/research\/wp-json\/microsoft-research\/v1\/researchers?person=Hao Cheng"},{"type":"user_nicename","value":"Mike Hang Wang","user_id":43951,"rest_url":"https:\/\/research.codeghost.online\/en-us\/research\/wp-json\/microsoft-research\/v1\/researchers?person=Mike Hang Wang"},{"type":"user_nicename","value":"Jianfeng Gao","user_id":32246,"rest_url":"https:\/\/research.codeghost.online\/en-us\/research\/wp-json\/microsoft-research\/v1\/researchers?person=Jianfeng Gao"},{"type":"text","value":"Sharon Li","user_id":0,"rest_url":false}],"msr_impact_theme":[],"msr_research_lab":[199565],"msr_event":[],"msr_group":[],"msr_project":[],"publication":[],"video":[],"msr-tool":[],"msr_publication_type":"misc","related_content":[],"_links":{"self":[{"href":"https:\/\/research.codeghost.online\/en-us\/research\/wp-json\/wp\/v2\/msr-research-item\/1180547","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/research.codeghost.online\/en-us\/research\/wp-json\/wp\/v2\/msr-research-item"}],"about":[{"href":"https:\/\/research.codeghost.online\/en-us\/research\/wp-json\/wp\/v2\/types\/msr-research-item"}],"version-history":[{"count":2,"href":"https:\/\/research.codeghost.online\/en-us\/research\/wp-json\/wp\/v2\/msr-research-item\/1180547\/revisions"}],"predecessor-version":[{"id":1180710,"href":"https:\/\/research.codeghost.online\/en-us\/research\/wp-json\/wp\/v2\/msr-research-item\/1180547\/revisions\/1180710"}],"wp:attachment":[{"href":"https:\/\/research.codeghost.online\/en-us\/research\/wp-json\/wp\/v2\/media?parent=1180547"}],"wp:term":[{"taxonomy":"msr-research-highlight","embeddable":true,"href":"https:\/\/research.codeghost.online\/en-us\/research\/wp-json\/wp\/v2\/msr-research-highlight?post=1180547"},{"taxonomy":"msr-research-area","embeddable":true,"href":"https:\/\/research.codeghost.online\/en-us\/research\/wp-json\/wp\/v2\/research-area?post=1180547"},{"taxonomy":"msr-publication-type","embeddable":true,"href":"https:\/\/research.codeghost.online\/en-us\/research\/wp-json\/wp\/v2\/msr-publication-type?post=1180547"},{"taxonomy":"msr-publisher","embeddable":true,"href":"https:\/\/research.codeghost.online\/en-us\/research\/wp-json\/wp\/v2\/msr-publisher?post=1180547"},{"taxonomy":"msr-publication-cta","embeddable":true,"href":"https:\/\/research.codeghost.online\/en-us\/research\/wp-json\/wp\/v2\/msr-publication-cta?post=1180547"},{"taxonomy":"msr-focus-area","embeddable":true,"href":"https:\/\/research.codeghost.online\/en-us\/research\/wp-json\/wp\/v2\/msr-focus-area?post=1180547"},{"taxonomy":"msr-locale","embeddable":true,"href":"https:\/\/research.codeghost.online\/en-us\/research\/wp-json\/wp\/v2\/msr-locale?post=1180547"},{"taxonomy":"msr-post-option","embeddable":true,"href":"https:\/\/research.codeghost.online\/en-us\/research\/wp-json\/wp\/v2\/msr-post-option?post=1180547"},{"taxonomy":"msr-field-of-study","embeddable":true,"href":"https:\/\/research.codeghost.online\/en-us\/research\/wp-json\/wp\/v2\/msr-field-of-study?post=1180547"},{"taxonomy":"msr-conference","embeddable":true,"href":"https:\/\/research.codeghost.online\/en-us\/research\/wp-json\/wp\/v2\/msr-conference?post=1180547"},{"taxonomy":"msr-journal","embeddable":true,"href":"https:\/\/research.codeghost.online\/en-us\/research\/wp-json\/wp\/v2\/msr-journal?post=1180547"},{"taxonomy":"msr-impact-theme","embeddable":true,"href":"https:\/\/research.codeghost.online\/en-us\/research\/wp-json\/wp\/v2\/msr-impact-theme?post=1180547"},{"taxonomy":"msr-pillar","embeddable":true,"href":"https:\/\/research.codeghost.online\/en-us\/research\/wp-json\/wp\/v2\/msr-pillar?post=1180547"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}