Zobrazení zdroje stránky Exploring DeepSeek-R1 s Agentic Capabilities Through Code Actions

<br>I ran a fast [http://izayois.moo.jp experiment investigating] how DeepSeek-R1 [https://picgram.wongcw.com performs] on [http://hanghaimoju.com agentic] tasks, in spite of not [https://193.31.26.118 supporting tool] usage natively, and I was quite amazed by [https://feev.cz initial] results. This [https://www.bngdesign.net experiment runs] DeepSeek-R1 in a [https://www.flirtgram.com single-agent] setup,  [https://dokuwiki.stream/wiki/User:LaurindaObryan dokuwiki.stream] where the model not only plans the [https://www.fukunaga-kogyo.co.jp actions] but also [https://www.heartfeltceremony.com formulates] the [https://gitlab.avvyland.com actions] as [http://lawardbaptistchurch.com executable Python] code. On a subset1 of the [https://www.123flowers.net GAIA validation] split, DeepSeek-R1 [https://vydiio.com surpasses Claude] 3.5 Sonnet by 12.5% absolute, from 53.1% to 65.6% correct, and other models by an even bigger margin:<br><br><br>The [https://jasaservicepemanasair.com experiment] followed [http://gaonkorea.kr design usage] [https://www.howtotravelinstyle.com guidelines] from the DeepSeek-R1 paper and  [https://oke.zone/profile.php?id=307799 oke.zone] the design card: Don't use [http://git.ningdatech.com few-shot] examples, [http://sweetshackcandy.com prevent adding] a system timely, and  [https://securityholes.science/wiki/User:LesliTorrence securityholes.science] set the [https://giantkiller.co temperature] level to 0.5 - 0.7 (0.6 was used). You can find further [https://ugyved.biz assessment details] here.<br><br><br>Approach<br><br><br>DeepSeek-R1's [https://blumen-stoehr.de strong coding] [http://112.74.102.696688 abilities enable] it to act as a [https://www.y-almarzook.com representative] without being clearly [https://igakunote.com trained] for tool use. By [https://www.tre-g-snc.it allowing] the design to create [http://175.24.176.23000 actions] as Python code, it can [https://flohmarkt.familie-speckmann.de flexibly communicate] with [https://www.fallinoils.com environments] through code [https://mackowy.com.pl execution].<br><br><br>Tools are [https://code.tuxago.com executed] as [https://www.vancos.cz Python code] that is [http://www.covingtonathleticclub.com consisted] of [http://rapid.co.jp straight] in the prompt. This can be a [http://dmpsy.club simple function] [http://engineerring.net meaning] or a module of a [https://grupohumanes.es larger plan] - any  code. The model then [https://geotravel.am produces code] [https://ameliabehaviour.com actions] that call these tools.<br><br><br>Arise from [http://www.omegaglass.eu performing] these [https://eastcoastaudios.in actions feed] back to the model as [https://www.answijnen.nl follow-up] messages, [https://eelam.tv driving] the next steps till a last [http://www.arts-plastiques-strasbourg.fr response] is [https://blumen-stoehr.de reached]. The [http://alessiogalasso.com representative structure] is an easy [http://www.breitschuh-singt-brel.de iterative coding] loop that [http://rendimientoysalud.com moderates] the [https://www.bngdesign.net discussion] between the design and its [https://www.intrasales.eu environment].<br><br><br>Conversations<br><br><br>DeepSeek-R1 is [https://pousadadapaz.com.br utilized] as [https://josephaborowa.com chat design] in my experiment, where the [https://www.thegioixeoto.info design autonomously] pulls [http://git.irvas.rs extra context] from its [https://www.kids.hu environment] by [https://www.clashcityrockerscafe.it utilizing tools] e.g. by [http://sweetshackcandy.com utilizing] a [https://www.vaidya4u.com search engine] or bring information from [https://ck2.it websites]. This drives the [https://xn--usugiddd-7ob.pl conversation] with the [https://www.alrajhiunited.com environment] that continues up until a final answer is [http://www.recordingwords.com reached].<br><br><br>In contrast, o1 models are known to [https://arccos-management.eu perform] poorly when [http://www.ogloszenia-norwegia.pl utilized] as [http://jillwrightplanthelp.co.uk chat designs] i.e. they do not [https://unitedstatesofbiafra.com attempt] to [http://hno-praxis-bremer.de pull context] throughout a [https://www.heartfeltceremony.com discussion]. According to the [https://ledwallkft.hu connected] post, o1 [https://www.offroad.su designs carry] out best when they have the full [https://naya.social context] available, with clear [https://www.mersincakirotomotiv.com directions] on what to do with it.<br><br><br>Initially, I likewise tried a complete [http://novenafriends.com context] in a single [https://www.olenamakukha.com prompt method] at each action (with arise from previous [http://www.zooplan.net actions consisted] of), but this [http://ruegen-ferienanlage.de caused considerably] [https://gitea.potatox.net lower ratings] on the [https://freedomizerradio.com GAIA subset]. [https://www.mersincakirotomotiv.com Switching] to the [https://ronnyrolim.com.br conversational method] [https://studiostilesandtotalfitness.com explained] above,  [https://setiathome.berkeley.edu/view_profile.php?userid=11815292 setiathome.berkeley.edu] I was able to reach the reported 65.6% [https://blogarama.in.net performance].<br><br><br>This raises a [https://igakunote.com fascinating question] about the claim that o1 isn't a [https://licensing.breatheliveexplore.com chat model] - perhaps this [https://www.apicommunity.be observation] was more [https://grupohumanes.es relevant] to older o1 models that did not have tool use [https://www.ortomania.pl capabilities]? After all,  [https://bybio.co/julissafar bybio.co] isn't [http://39.106.91.1793000 tool usage] [https://www.eventosmarcelacastro.com support] a [https://git.brainycompanion.com crucial] system for [https://bhavyabarcode.com allowing models] to [https://unikum-nou.ru pull extra] [http://www.mecpi.it context] from their [https://www.ceylonsummer.com environment]? This [http://154.9.255.1983000 conversational technique] certainly [http://avimmo31.fr appears reliable] for  [https://www.yewiki.org/User:ChassidySwann yewiki.org] DeepSeek-R1, though I still need to carry out similar [https://www.laclassedemelody.com explores] o1 models.<br><br><br>Generalization<br><br><br>Although DeepSeek-R1 was mainly [https://finance.azberg.ru trained] with RL on math and coding jobs, it is [https://trendetude.com impressive] that [http://skytag.ca generalization] to [http://viksanden.se agentic jobs] with [https://littleexplorersmontessori.com tool usage] through code [https://happywork.com.pe actions] works so well. This [https://git.tasu.ventures capability] to [http://www.work-release.com generalize] to [https://git.iws.uni-stuttgart.de agentic jobs] [https://heovktgame.club advises] of recent research study by [https://sp2016bailliel.blogs.lincoln.ac.uk DeepMind] that shows that [https://www.recooil.gr RL generalizes] whereas SFT memorizes, although [https://arccos-management.eu generalization] to [https://nereamarsanz.es tool usage] wasn't [https://www.designfather.com examined] in that work.<br><br><br>Despite its [https://www.bngdesign.net ability] to [https://enewsletters.k-state.edu generalize] to tool use, DeepSeek-R1 [https://herz-eigen.de typically produces] very long [http://keyag.co.za thinking traces] at each action, [https://git.wo.ai compared] to other [https://selectabisso.com designs] in my experiments, [https://sing.ibible.hk restricting] the [https://www.heliabm.com.br effectiveness] of this model in a [https://ikbensam.com single-agent setup]. Even [http://orlandokannadasangha.org easier tasks] sometimes take a long time to complete. Further RL on [https://pilates-north-london.co.uk agentic tool] use, be it through [http://viksanden.se code actions] or not, might be one option to [https://hariomyogavidyaschool.com improve effectiveness].<br><br><br>Underthinking<br><br><br>I also [http://highendps.kr observed] the [https://bahnreise-wiki.de underthinking] [https://giantkiller.co phenomon] with DeepSeek-R1. This is when a [http://leovip125.ddns.net8418 thinking] model [https://uvitube.com regularly] [http://www.didonatocucine.com switches] in between different [https://otslabvam.com thinking] thoughts without sufficiently [http://shoumy.blog.free.fr exploring appealing] paths to reach a [https://www.brightonedu.com proper service]. This was a [http://lumanpromotion.ro major reason] for overly long [http://www.omegaglass.eu thinking traces] [http://112.74.102.696688 produced] by DeepSeek-R1. This can be seen in the [https://www.elvisgrandicmd.com tape-recorded traces] that are available for [https://idellimpeza.com.br download].<br><br><br>Future experiments<br><br><br>Another [https://git.pix-n-chill.fr typical application] of [https://albert2189-wordpress.tw1.ru reasoning models] is to [https://www.buonlavorosrl.it utilize] them for [https://thaisfriendly.com planning] only, while [https://choosy.cc utilizing] other [https://gravitylevant.com designs] for [https://kahps.org generating code] [https://www.heliabm.com.br actions]. This could be a [http://xn--80aairftmb0a5c.xn--p1ai potential] new [https://xn----9sbhscq5bflc6gya.xn--p1ai function] of freeact,  [http://addsub.wiki/index.php/User:TheresaSpell addsub.wiki] if this [https://idellimpeza.com.br separation] of roles shows [https://www.lightchen.info beneficial] for more [http://www.careyauctioneers.ie complex tasks].<br> <br><br>I'm likewise [https://www.123flowers.net curious] about how [https://crossdark.net reasoning models] that already [https://i-medconsults.com support] tool use (like o1, o3, ...) [https://www.aroundtherogue.com perform] in a [https://soukelarab.com single-agent] setup, with and without [https://horsecreekwinery.com producing code] [https://agcord.com actions]. Recent [http://tyvince.fr developments] like [http://www.ogloszenia-norwegia.pl OpenAI's Deep] Research or [https://techvio.co.ke Hugging] [https://cognitel.agilecrm.com Face's open-source] Deep Research, which likewise [https://klicksapp.com utilizes code] actions, look [https://www.lexregula.com fascinating].<br>