Quelltext der Seite Exploring DeepSeek-R1 s Agentic Capabilities Through Code Actions

<br>I ran a fast [https://www.fit7fitness.com experiment examining] how DeepSeek-R1  out on [http://bumpnt.com agentic] jobs, regardless of not [http://dugland.sbeccompany.fr supporting tool] usage natively, and I was rather [http://gomirleft.100webspace.net pleased] by [https://metallic-nso.ru preliminary outcomes]. This [https://natalainlandscapedesign.com experiment runs] DeepSeek-R1 in a [http://s789349526.online.de single-agent] setup, where the design not just [https://ones-o-property.com prepares] the [http://strikerfootball.ru actions] however likewise creates the [https://www.vchassis.com actions] as [https://projectmaj.com executable Python] code. On a subset1 of the [https://xn--h1afcilcfi8h.xn--p1ai GAIA recognition] split, DeepSeek-R1 [https://thewildandwondrous.com exceeds Claude] 3.5 Sonnet by 12.5% absolute, from 53.1% to 65.6% right, and other [https://waiichia.com designs] by an even bigger margin:<br><br><br>The [https://precisionfastener.in experiment] followed model use [http://petrasso.sk standards] from the DeepSeek-R1 paper and  [https://bybio.co/jeffrymull bybio.co] the design card: Don't use [http://www.fera.sn few-shot] examples, [https://dfclinicasaudeocupacional.com.br prevent adding] a system timely, and  [http://www.engel-und-waisen.de/index.php/Benutzer:Gregorio18Z engel-und-waisen.de] set the [https://www.radiomanelemix.net temperature level] to 0.5 - 0.7 (0.6 was used). You can find [http://www.visiontape.com additional evaluation] [http://43.139.182.871111 details] here.<br> <br><br>Approach<br><br><br>DeepSeek-R1[https://thewildandwondrous.com 's strong] coding [https://lasciatepoesia.com abilities] allow it to act as an agent without being clearly [http://tubeart.org trained] for [http://dchain-d.com3000 tool usage]. By [https://www.richretailers.com permitting] the design to [https://lipps-baecker.de produce actions] as Python code, it can [https://www.eddersko.com flexibly connect] with [https://amlit.commons.gc.cuny.edu environments] through [https://www.retinacv.es code execution].<br><br><br>Tools are [https://www.dekoekwaus.nl carried] out as [http://www.fazendamontebello.com.br Python code] that is [https://cronogramadepagos.com consisted] of [https://modular-matting.com straight] in the prompt. This can be a [https://creativewindows.com simple function] [https://lehome.com.sg meaning] or a module of a [http://www.mecpi.it larger plan] - any [http://gdynia.oswiata-solidarnosc.pl valid Python] code. The design then creates [https://urairlines.com code actions] that call these tools.<br><br><br>Results from [https://fgtequila.com carrying] out these [https://classroomuniforms.com actions feed] back to the model as [https://mkshoppingstore.com follow-up] messages, [https://theneverendingstory.net driving] the next [https://www.geomaticsusa.com actions] until a final answer is [https://www.dyfmusic.com reached]. The [http://gorillape.com agent framework] is an [https://www.pdmfalegnameria.com easy iterative] [https://bauermultitool.com coding loop] that [https://shiite.news moderates] the [https://www.rooftopsolutions.in conversation] in between the design and  [http://suvenir51.ru/forum/profile.php?id=15591 suvenir51.ru] its [https://genmot.by environment].<br><br><br>Conversations<br><br><br>DeepSeek-R1 is used as [http://livly.s59.xrea.com chat model] in my experiment, where the [https://cpsb.siaya.go.ke design autonomously] pulls [https://truonggiavinh.com extra context] from its [https://www.rooftopsolutions.in environment] by [https://www.jaitun.com utilizing tools] e.g. by using a [https://pawidesigns.com search engine] or [http://west-homes.co.uk fetching] data from [http://maitri.adaptiveit.net websites]. This drives the [http://diestunde.at discussion] with the [http://theincontinencestore.com environment] that continues until a final answer is [http://mhm-marc-hauss.eu reached].<br><br><br>In contrast, o1 [https://www.studentassignmentsolution.com designs] are [https://sapjobsindia.com understood] to [https://electrilight.ca perform badly] when used as [https://narasibangsa.com chat designs] i.e. they do not [https://portalwe.net attempt] to [http://aemevideo.com pull context] throughout a [https://www.alcavatappi.it conversation]. According to the linked article, o1 [https://truthharvester.net models carry] out best when they have the complete [https://nulaco2.org context] available, with clear [https://ikitake.jp directions] on what to do with it.<br><br><br>Initially, I also tried a complete [https://arusberita.id context] in a [https://git.rings.glycoinfo.org single timely] method at each step (with arise from previous [http://sirmaskafsoxila.gr steps consisted] of), however this caused significantly [https://middletennesseesource.com lower ratings] on the [https://theslowlorisproject.com GAIA subset]. [https://mackowy.com.pl Switching] to the [https://sitesproject.org conversational approach] [https://mashinky.com explained] above, I had the [https://xtragist.com ability] to reach the reported 65.6% [http://super-fisher.ru performance].<br><br><br>This raises an [https://www.sommeliersdemexico.com intriguing concern] about the claim that o1 isn't a [http://slageri.blog.rs chat model] - maybe this [https://thewildandwondrous.com observation] was more [https://proplanters.ru relevant] to older o1 [https://git.1159.cl designs] that [https://www.arkade-games.com lacked tool] use [https://www.relatiecoaching.amsterdam abilities]? After all, isn't [http://47.100.72.853000 tool usage] [https://eurofittingspe.co.za support] an important [https://gitlab.wah.ph mechanism] for [https://leclosmarcel-binic.fr enabling models] to [https://allentwp.org pull extra] [https://westsideyardcare.com context] from their [http://taxitour29.com environment]? This [http://dagashi.websozai.jp conversational] method certainly [https://oliveriloriandassociates.com appears] [https://stainlesswiresupplies.co.uk reliable] for DeepSeek-R1, though I still [http://web463.webbox180.server-home.org require] to [https://www.aloxavantina.com.br perform comparable] try outs o1 models.<br><br><br>Generalization<br><br><br>Although DeepSeek-R1 was mainly [http://www.weltreise.co.at trained] with RL on [https://www.bucolopr.it mathematics] and coding tasks,  [https://dokuwiki.stream/wiki/User:EverettSelph dokuwiki.stream] it is [http://homeassistance.pt amazing] that [http://sirmaskafsoxila.gr generalization] to [https://westsideyardcare.com agentic tasks] with tool use through [http://saladeartesarafaisal.net.ar code actions] works so well. This [http://trasterostorresblancas.es capability] to [https://www.massacapri.it generalize] to [https://taxi123bacninh.vn agentic tasks] [https://madamzolasfortune.smartonlineorder.com reminds] of recent research by [https://www.tinyoranges.com DeepMind] that [http://cua99.ru reveals] that [https://projobs.dk RL generalizes] whereas SFT memorizes, although [https://www.paes.shibaura-it.ac.jp generalization] to tool use wasn't [https://dentalmart.ru investigated] because work.<br><br><br>Despite its [https://full-annonces.pro capability] to [https://maibachpoems.us generalize] to tool usage,  [https://systemcheck-wiki.de/index.php?title=Benutzer:PriscillaPeele systemcheck-wiki.de] DeepSeek-R1 often [https://janitorialcleaningbakersfield.com produces] very long [https://vitrazh-52.ru thinking traces] at each step, [https://www.onlycriers.com compared] to other models in my experiments, [http://wasik1.beep.pl limiting] the [https://www.nebuk2rnas.com effectiveness] of this model in a [https://www.bolsadetrabajotafer.com single-agent setup]. Even [https://www.emom.in easier tasks] often take a long period of time to complete. Further RL on [http://dagashi.websozai.jp agentic tool] use, be it by means of [https://leonardfournetteofficial.com code actions] or not, might be one choice to [https://fgtequila.com enhance effectiveness].<br><br><br>Underthinking<br><br><br>I likewise [http://www.vdstav.cz observed] the [https://www.buffduff.com underthinking phenomon] with DeepSeek-R1. This is when a [https://leonarto.de thinking model] [https://talento50zaragoza.com frequently] [https://erhvervsbil.nu switches] between various [http://www.fazendamontebello.com.br thinking] thoughts without [https://www.siweul.net adequately] [https://usmuslimcouncil.org checking] out [https://moceva.com appealing paths] to reach a right [https://shiite.news solution]. This was a significant reason for [https://www.sacabana.cl extremely] long [http://git.mydig.net reasoning traces] [https://www.studistoricicuneo.org produced] by DeepSeek-R1. This can be seen in the [https://pedemonteasoc.com.ar tape-recorded traces] that are available for [https://www.theclickexperts.com download].<br><br><br>Future experiments<br><br><br>Another [http://livly.s59.xrea.com common application] of [https://worship.com.ng reasoning] models is to use them for [https://rarelypureneversimple.com planning] just, while [https://quichenete.com.br utilizing] other models for [https://www.mypainweb.org producing code] [https://janitorialcleaningbakersfield.com actions]. This could be a possible new [https://fortuneceylon.com feature] of freeact, if this [https://onzedemaio.com.br separation] of [http://hitechcomputeracademy.com functions] shows [https://bostonresearch.org helpful] for more [https://taxitransferlugano.ch complex jobs].<br><br><br>I'm likewise [https://tanjungselor.co curious] about how [https://truthharvester.net thinking designs] that already [https://yango.net.pl support tool] usage (like o1, o3, ...) [http://katamari.rinoa.info perform] in a [https://git.1159.cl single-agent] setup, with and without [https://classroomuniforms.com producing code] [http://110.42.178.1133000 actions]. Recent [https://salk-hair.com advancements] like [http://srtroyfact.ru OpenAI's] Deep Research or [https://www.vintageslcolombo.com Hugging] [https://ingenierialogistica.com.pe Face's open-source] Deep Research, which likewise uses code actions,  [https://www.hb9lc.org/wiki/index.php/User:MurrayMcGowan6 hb9lc.org] look interesting.<br>