{"about":{"non_affiliation":"Code with Papers and Syntology are not affiliated with, endorsed by, or sponsored by Papers with Code, Meta, or the pwc-archive mirror.","site":"https://codewithpapers.app","licence":"CC BY-SA 4.0","licence_url":"https://creativecommons.org/licenses/by-sa/4.0/legalcode","attribution":"https://codewithpapers.app/attribution","modified":"archive material modified by Syntology; see the attribution page","syntology":{"site":"https://syntology.ai","developers":"https://syntology.ai/developers","mcp":{"server":"https://syntology.ai/mcp","transport":"streamable-http","server_card":"https://syntology.ai/.well-known/mcp/server-card.json","auth":{"type":"trial token, no account","trial_token":"https://syntology.ai/api/oauth/trial/token","method":"POST","docs":"https://syntology.ai/developers"}},"have":"https://syntology.ai/api/graph/have?x=<method, arXiv id or title> (free, answers coverage only)","paper_base":"https://syntology.ai/paper/","atlas_base":"https://app.syntology.ai/?focus="},"machine_readable":[{"url":"https://codewithpapers.app/llms.txt","what":"the machine catalog: every machine-readable file, counted"},{"url":"https://codewithpapers.app/index/manifest.json","what":"paper-to-code index by arXiv id, with Syntology's counts"},{"url":"https://codewithpapers.app/search/manifest.json","what":"site search index (titles, authors) and its files"},{"url":"https://codewithpapers.app/download","what":"bulk files: Syntology's layer, described there"},{"url":"https://codewithpapers.app/build_manifest.json","what":"the build record: inputs, counts, exclusions, probes"}]},"url":"/paper/is-reinforcement-learning-not-for-natural","title":"Is Reinforcement Learning (Not) for Natural Language Processing: Benchmarks, Baselines, and Building Blocks for Natural Language Policy Optimization","arxiv_id":"2210.01241","date":"2022-10-03","proceeding":null,"authors":["Rajkumar Ramamurthy","Prithviraj Ammanabrolu","Kianté Brantley","Jack Hessel","Rafet Sifa","Christian Bauckhage","Hannaneh Hajishirzi","Yejin Choi"],"abstract":"We tackle the problem of aligning pre-trained large language models (LMs) with human preferences. If we view text generation as a sequential decision-making problem, reinforcement learning (RL) appears to be a natural conceptual framework. However, using RL for LM-based generation faces empirical challenges, including training instability due to the combinatorial action space, as well as a lack of open-source libraries and benchmarks customized for LM alignment. Thus, a question rises in the research community: is RL a practical paradigm for NLP? To help answer this, we first introduce an open-source modular library, RL4LMs (Reinforcement Learning for Language Models), for optimizing language generators with RL. The library consists of on-policy RL algorithms that can be used to train any encoder or encoder-decoder LM in the HuggingFace library (Wolf et al. 2020) with an arbitrary reward function. Next, we present the GRUE (General Reinforced-language Understanding Evaluation) benchmark, a set of 6 language generation tasks which are supervised not by target strings, but by reward functions which capture automated measures of human preference.GRUE is the first leaderboard-style evaluation of RL algorithms for NLP tasks. Finally, we introduce an easy-to-use, performant RL algorithm, NLPO (Natural Language Policy Optimization)} that learns to effectively reduce the combinatorial action space in language generation. We show 1) that RL techniques are generally better than supervised methods at aligning LMs to human preferences; and 2) that NLPO exhibits greater stability and performance than previous policy gradient methods (e.g., PPO (Schulman et al. 2017)), based on both automatic and human evaluations.","url_abs":"https://arxiv.org/abs/2210.01241v3","url_pdf":"https://arxiv.org/pdf/2210.01241v3.pdf","source":{"archive":"pwc-archive (Hugging Face), CC BY-SA 4.0","snapshot":"2025-07-28","licence_url":"https://creativecommons.org/licenses/by-sa/4.0/legalcode","row_kind":"abstracts"},"code_links":[{"paper_slug":"is-reinforcement-learning-not-for-natural","repo_url":"https://github.com/allenai/rl4lms","is_official":1,"mentioned_in_paper":1,"mentioned_in_github":1,"framework":"pytorch","reach":{"status":"ok","spdx":"Apache-2.0"}},{"paper_slug":"is-reinforcement-learning-not-for-natural","repo_url":"https://github.com/debjitpaul/refiner","is_official":0,"mentioned_in_paper":0,"mentioned_in_github":1,"framework":"pytorch","reach":{"status":"ok","spdx":"Apache-2.0"}},{"paper_slug":"is-reinforcement-learning-not-for-natural","repo_url":"https://github.com/tedmoskovitz/constrainedrl4lms","is_official":0,"mentioned_in_paper":0,"mentioned_in_github":1,"framework":"pytorch","reach":{"status":"ok"}}],"tasks":[{"task_slug":"decision-making","task_name":"Decision Making"},{"task_slug":"policy-gradient-methods","task_name":"Policy Gradient Methods"},{"task_slug":"reinforcement-learning-1","task_name":"Reinforcement Learning (RL)"},{"task_slug":"sequential-decision-making","task_name":"Sequential Decision Making"},{"task_slug":"text-generation","task_name":"Text Generation"},{"task_slug":"reinforcement-learning-2","task_name":"reinforcement-learning"}],"methods":[{"method_slug":"entropy-regularization","method_name":"Entropy Regularization"},{"method_slug":null,"method_name":"Library"},{"method_slug":"ppo","method_name":"PPO"}],"datasets_introduced":[],"methods_introduced":[],"results":[],"syntology":{"syntology_url":"https://syntology.ai/paper/2210.01241","atlas_url":"https://app.syntology.ai/?focus=2210.01241","mcp":null,"developers":"https://syntology.ai/developers","held":true,"reason":"arxiv_id_in_atlas","checked_against":{"atlas_ids":549449,"atlas_ids_sha256":"afd72399dafd9906e9062927f70655952385a3d75488558e1f0484e950ab09b7","read_at":"2026-10-05T01:46:21Z"},"agent_calls":[{"tool":"get_reference_implementation","arguments":{"name":"NLPO"},"call":"get_reference_implementation(name='NLPO')","min_level":{"quoted_from_server_card":["It REFUSES rather than serving below min_level, and the refusal names the highest level that exists","so lowering min_level is a deliberate, informed choice, not a guess."],"default_on_card":2,"server_card":"https://syntology.ai/.well-known/mcp/server-card.json","card_sha256":"479785a84658dd682a11b10237b3620d651d7f87799f5dc7add970aaa83f6d39","this_implementation_level":3}},{"tool":"get_citation_path","arguments":{"paper_1":"2210.01241"},"arguments_in":{"paper_2":"another paper's arXiv id or title"},"call":"get_citation_path(paper_1=\"2210.01241\", paper_2=\"…\")"},{"tool":"get_concepts_for_paper","arguments":{"arxiv_id":"2210.01241"},"call":"get_concepts_for_paper(arxiv_id=\"2210.01241\")"}],"reference_implementation":[{"method_name":"NLPO","assembly_sha256":"8396c6d86828eae4","source":"generated from the paper text (not the authors' code)","level":3,"level_is_quoted_from_server_card":[{"level":"V2","quoted":["two implementations GENERATED from the paper's signature and spec agree numerically (1e-6) on fixture inputs","neither of the two ever read the authors' own code, so V2 measures how unambiguously the paper specifies the method, not fidelity to the authors' implementation (measured 2026-09-11, commit 75eaa77: cross_checked_by is a second LLM on 2,628 of 2,831 served samples)"]},{"level":"V3","quoted":["paper-derived property tests pass on both of those same two generations."]}],"server_card":"https://syntology.ai/.well-known/mcp/server-card.json","card_sha256":"479785a84658dd682a11b10237b3620d651d7f87799f5dc7add970aaa83f6d39","v3_tests_passed_both_sides":0,"level_3_note":"level 3 is recorded, but its report lists no property test that passed on both","no_rung":"No rung compares this code against the authors' implementation","no_rung_quoted_from_server_card":"No rung on this ladder compares the served code against the authors' implementation, so treat every level as evidence about runnability and internal agreement, not about matching the paper's results.","retrieve_with":"get_reference_implementation(name='NLPO')","have_check":{"url":"https://syntology.ai/api/graph/have?x=2210.01241","call":"have('2210.01241')","in_build_sample":false}}],"reference_implementation_is":"Syntology's reference implementation of this paper's method: code Syntology generated from the paper text (not the authors' code), at the level its own end-to-end run recorded (the assembly's own level, never averaged); the level's meaning is quoted from Syntology's server card, and nothing here says the code is correct or reproduces the paper's results","by_repo":[{"repo":"allenai/rl4lms","kind":"official","samples":0,"ran":0,"instrument":0,"unverified":0,"harvested_for_other_papers":1,"harvested_for":["2305.08844"],"harvested_for_calls":[{"arxiv_id":"2305.08844","tool":"get_harvested_code_for_paper","arguments":{"arxiv_id":"2305.08844"},"call":"get_harvested_code_for_paper(arxiv_id=\"2305.08844\")"}]},{"repo":"debjitpaul/refiner","kind":"listed","samples":0,"ran":0,"instrument":0,"unverified":0,"harvested_for_other_papers":1,"harvested_for":["2304.01904"],"harvested_for_calls":[{"arxiv_id":"2304.01904","tool":"get_harvested_code_for_paper","arguments":{"arxiv_id":"2304.01904"},"call":"get_harvested_code_for_paper(arxiv_id=\"2304.01904\")"}]},{"repo":"tedmoskovitz/constrainedrl4lms","kind":"listed","samples":0,"ran":0,"instrument":0,"unverified":0,"harvested_for_other_papers":1,"harvested_for":["2310.04373"],"harvested_for_calls":[{"arxiv_id":"2310.04373","tool":"get_harvested_code_for_paper","arguments":{"arxiv_id":"2310.04373"},"call":"get_harvested_code_for_paper(arxiv_id=\"2310.04373\")"}]}],"by_repo_harvested_for_is":"harvested_for: the arXiv ids (as Syntology's graph holds them) of every paper whose samples Syntology's graph links to this repository, in arXiv-id order, complete; the page names the first 3 and says how many more. A link only: it does not say the papers are the same work, that one copied the other, or that the code implements this paper's method; harvested_for_calls gives the get_harvested_code_for_paper call for each arXiv-shaped id among them","by_repo_is":"one row per repository in the archive's code links, Syntology's graph links or the samples; samples 0 means no sample is linked to this paper, a statement about Syntology's coverage, not about the repository: with harvested_for_other_papers present, Syntology harvested the repository for that many other papers, and without it, Syntology harvested nothing from it; placed_by_identical_code counts samples whose link records no repository, placed here because identical code was harvested from this repository for another paper; ran counts executions on a synthesized input and instrument counts failures of Syntology's instrument, not of the code","assemblies":[{"assembly_sha256":"8396c6d86828eae4","source_kind":"generated","method_name":"NLPO","realizes_method":"NLPO","graph_method":"NLPO","realizes_severed":false,"claim":"Compute the PPO-Clip objective for a batch of timesteps. Input log_probs_new is a 1D array of shape (N,) containing log probabilities of actions under the current policy. Input log_probs_old is a 1D array of shape (N,) containing log probabilities of the same actions under the old policy. Input adva","centrality":"contribution","entry":"ppo_clip_objective","entry_sha256":"0385a2f7384cb25949685505b639ecab100112660126e6236897f62c04344fc3","n_bricks":1,"bricks":[{"entry":"ppo_clip_objective","code_sha256":"0385a2f7384cb25949685505b639ecab100112660126e6236897f62c04344fc3","role":"entry","order":0,"level":3,"status":"ran_honours"}],"ran_individually":1,"ran_individually_constructed":null,"ran_individually_by_status":{"ran_honours":1},"end_to_end":{"ran":true,"level":3,"against_independent_peer":true,"n_peers":1,"v3_tests_passed_both_sides":0,"means":"this assembly is a single routine, so the entry brick's verification IS the assembly's: it executed and two family-independent implementations agreed"},"level_lowered":null,"claim_match":"IMPLEMENTS","claim_match_rule":"single_judge","claim_match_basis":"construction","claim_match_judges":["construction:spec_drafted_for_this_method"],"label_verdict":"matches","label_verdict_judged":false,"suggested_method":null,"batch_at":"2026-09-07T20:44:12+00:00","sibling_group":"9e0821d5f8d8","n_siblings":0,"retention":"keep","provenance":"assembly:generated_reference_impl_v1","attached_at":"2026-09-07T20:44:12+00:00","attached_provenance":"assembly:generated_reference_impl_v1"}],"assembly_sibling_groups":[{"group":"9e0821d5f8d8","realizes_method":"NLPO","source_kind":"generated","batch_at":"2026-09-07T20:44:12+00:00","members":["8396c6d86828eae4"]}],"assemblies_is":{"about":"What Syntology's graph records as this paper's claimed method, lifted into code: each assembly names its bricks (functions and classes), each with its own role and its own verification level. An assembly is not a check that the code implements the method correctly or reproduces the paper's results.","harvested":"assembled from code harvested from a repository linked to the paper","generated":"generated from the paper text, not the repository's code","end_to_end":"for a harvested assembly, the assembly has not been run end-to-end; for a generated one, end_to_end.level is the assembly's own level and against_independent_peer says whether it ran against a peer: an independent peer is a second implementation generated separately from the same paper text; neither implementation read the authors' code. At level 2 the two agreed numerically on the fixture's inputs; level 3 is recorded when property tests derived from the paper also passed on both, and each level 3 card prints how many its report lists","bricks_level":"each brick's own verification level in Syntology's graph (level 0: no recorded run), listed brick by brick and never combined into one level","ran_individually":"a COUNT of bricks that ran on their own on a fixture (split by status in ran_individually_by_status), not a level; instrument statuses are failures of Syntology's instrument, not of the code","siblings":"assemblies lifted from this paper in the same batch for the same method: alternative readings kept side by side; none replaces or supersedes another","judged":"Where a model read the code against the paper's claim, each assembly's label verdict is one model's judgement, kept in this page's JSON twin; it is not calibrated, so no card prints it. A generated assembly's label was set when its code was generated for the method, and no model judged it; its card prints it as \"by construction\" or \"not judged by a model\"","level_lowered":"a generated assembly whose level Syntology's graph lowered, with the level it had, the recorded reason and when","retention":"Syntology's spending triage for further verification, not a verdict","claim_match":"the claim_match vote: IMPLEMENTS, COMPONENT or UNCLEAR. Where claim_match_basis is \"judge:<model>\", it is one model's judgement of the assembly's label against the paper's claim (claim_match_rule single_judge, the model in claim_match_judges): not a verification and not a run. Where it is \"construction\", the value was set when the code was generated for this method and no model judged it, whatever claim_match_rule says. The page prints no claim match for any assembly","label_verdict":"whether the assembly's code is the method it is labelled with: matches, cannot_tell, mislabelled (suggested_method names what it looks like instead) or not_the_method; label_verdict_judged is true only when a model read the code (claim_match_basis \"judge:<model>\")","label_verdict_judged":"where true (claim_match_basis \"judge:<model>\"), label_verdict is a single model's judgement; not calibrated; not printed on the page; where false, it was set by construction, and the page prints it unless the node is shared with another paper (node_properties_written_for)","claim":"the claim text stored on the assembly node: for a harvested assembly, what the paper claims the code implements, as the lifting model wrote it; for a generated one, the specification the code was generated from. Not verified text","centrality":"how central the lifted method is to the paper, as the lifting step recorded it: contribution, component or incidental","v3_tests_passed_both_sides":"end_to_end.v3_tests_passed_both_sides: the number of property tests derived from the paper that passed on both implementations, as the attached report lists them; a level 3 card with 0 says its report lists none","ran_individually_constructed":"ran_individually_constructed: of the bricks counted in ran_individually, how many constructed an object rather than computing a result (Syntology's RAN record's rule, read from the graph-layer sample with the brick's code_sha256); a part of ran_individually, never subtracted from it and never a failure; null where a brick that ran has no sample in the graph layer this build read, so the figure was not counted"},"assemblies_call":{"tool":"assembly","arguments":{"arxiv_id":"2210.01241"},"call":"assembly(arxiv_id=\"2210.01241\")"}},"code_links_note":null,"arxiv_metadata":null,"syntology_extracted_results":null,"record_sha256":"59662a9e9b0385a953eb70abdcd46d35a7d084660918dd5984f9e10b98b4f13e","record_changed_at":"2026-09-28","record_changed_at_basis":"first_hashed"}