Dzokera kuNhau
ChengetedzoAI Understanding muchidimbu

Kudzidza Kunowana Multi-Agent AI Kuchengetedzeka kunoenderana neProvenance, Kwete Chete Mitemo

Iyo itsva POLIS preprint inoshuma kuti chengetedzo inogoneka inogona kutadza kana vachivimba neinochinjika policy state pachinzvimbo chechiremera chakaigadzira, ukuwo ichiratidza kuti sei nzira dzekudzoreredza dzichikosha mushure mekuvhara.

6 min readRead the primary source
Primary-source documentKwakanyorwa
Muparidzi
POLIS multi-agent AI safety paper on arXiv
Source link
arxiv.orghttps://arxiv.org/abs/2608.09828
Source type
Gwaro rekutanga - chiziviso chepamutemo, bepa, faira, kana peji rebato rekutanga ratinoverenga zvakananga.
ContextNzwisisa izvi mumasekonzi makumi matanhatu

Tanga pano

Matemu akakosha

AI Kuchengetedza
Munda wakatarisana nekudzikisira maitiro anokuvadza, kutadza, uye njodzi yekushandisa zvisizvo muAI masisitimu.
Benchmark
Muedzo wakamisikidzwa kana dhatabheti rinoshandiswa kuyera nekuenzanisa kuita kwemuenzaniso.
Chikamu
Musiyano wekupinza unoshandiswa nemodhi kuita fungidziro.
Zviedze iwe pachakoAI Safety Quiz

Chii chaitika

A new arXiv preprint kubva kuPOLIS chirongwa chekutsvagisa chinoedza mubvunzo unonyanya kukosha sezvo vamiririri veAI vanotumira mabasa, kufambisa ruzivo, uye kufonera maturusi akamiririra masangano: mutemo wekuchengetedza unoshanda here kana sisitimu isingakwanisi kutaura kuti chiremera chiri shure kwemutemo iwoyo chakabva kupi? Mune yakaomeswa nechando 5,280-episode yekudzidza suite, vanyori vakasiyana mitemo yehutongi, nyika ine masimba yakabvunzurudzwa nemurindi anogona kuuraya, kukwezva kwemukati kudonha, uye kana yakavharika mafambiro angaenderera. Mhedzisiro iyi ongororo yeanodzorwa akawanda-agent workflows, kwete humbowo hwekuti chigadzirwa chakatumirwa chatambura nechiitiko.

Kuedza kukuru kwevamiriri kunobata mhuri ina dzemhando, ine yakanangwa yepamusoro-kukonana kuongororwa ichiwedzera mamwe matatu emodheni ekupedzisira. MaAgents anoshanda munzvimbo dzakarongeka umo deterministic system inosarudza kuti ndezvipi zviito zvinoitwa uye kana nyika inoguma ichityora mutemo. Dhizaini iyi inoita kuti vaongorori vaparadzanise chikumbiro chakarambidzwa kubva pakutyorwa kwakaitika, vobva vaenzanisa chirevo chakadzama chebumbiro nevarindi vanogona kuuraya vanopindira pamuganho wechiito.

Bepanhau rinoti kurudziro yebumbiro yakaburitsa zero dzakatyorwa muzvikamu mazana matatu nemakumi masere neina, pamwe ne365 dzakatemerwa. Murindi wepanzvimbo anogona kutongerwa akabvunzurudza nyika inooneka yemutemo akaburitsa kutyorwa kwemakumi maviri nembiri muhukuru hwakafanana, kana 5.7%. A provenance-aware executable guard akaburitsawo zero akatyorwa, asi yakavharira 51 yakarambidzwa kuedza; 44 yezvikamu izvozvo zvakazopera zvakachengeteka. Nhamba idzi dzinotsanangura mamiriro echiedzwa ichi, kwete mwero wekuchengetedzwa kwepasi rose wemhando dzemitauro.

Musiyano wakapinza wakaonekwa mune inofananidzwa nepolicy-laundering scenarios. Shanduko yakajairika yakashandura mutemo waionekwa nemurindi wepanzvimbo ichisiya masimba ekutanga asina kuchinjwa. Pasi pemamiriro ezvinhu akadaro, muchengeti wenharaunda akabvuma kutyorwa muzvikamu makumi maviri nezviviri zvemakumi mapfumbamwe nematanhatu, ukuwo kuchengetedzwa kwemutemo hakuna kubvuma, nebepa rekuzivisa p = 4.77 × 10⁻⁷. Iyo meshini yakamanikana pane chirevo chekuti vamiririri vanogona kupfuura chero chevarindi: chiratidzo chekuti nyika inotemerwa layer inovimba inogona kushandura mhedzisiro.

Chiyedzo chakasiyana chekugovera zviwanikwa chakawana kuti kuburitsa kukosha kwenhamba yekapu yakangofanana kwakachinja zvikumbiro zveagent. Vanyori vanoshandisa mhedzisiro iyoyo kupokana kuti kutyora kwekupedzisira kuverenga kunogona kuvanza nzira dzakasiyana: mutemo, rekodhi rekodhi, zvinokurudzira zvakagadzirwa nekufambiswa kwebasa, uye nzira inowanikwa mushure mekupindira nyaya dzese. Iro bepa rine mapeji gumi nemanomwe ane nhamba shanu, uye vanyori varo vanoti kodhi uye reproducibility artifacts inowanikwa munzvimbo yeruzhinji yePOLIS repository.

Kwakabva mashoko: POLIS multi-agent AI safety paper on arXiv ↗

Nei zvichikosha

Mharidzo yezvidzidzo ndeyekuti mvumo yeagentic system inofanira kuva yakasimba kudarika zvinyorwa zvitsva zvinomiririra mutemo. Kana mumiririri achigona kushandura data, kugovera basa, kana kuyambuka muganhu wesangano, gadhi anongotarisa mamiriro arikuoneka chete anogona kurasikirwa nenhoroondo inodiwa kuti asarudze kana chiito chichiri kubvumidzwa.

Musiyano iwoyo une basa kune masisitimu anobata zvinyorwa, marekodhi, marekodhi evatengi, kana maturusi akagovaniswa. Faera rinogona kukopwa, kupfupikiswa, kuputirwa mune chimwe chinhu, kana kupfuudzwa pakati pevamiririri nepo mavambo ayo uye zvirambidzo zvichiramba zvisina kuchinjwa. Kana chimwe chikamu chekupedzisira chichivimba chete neinochinjika label, neimwe nzira shanduko yakajairika inogona kuita kuti kutamisa kunorambidzwa kutaridzike kwakatenderwa. Provenance saka haisi yekungoongorora ficha; mune pepa rekuchenesa mamiriro, chikamu chesarudzo yekumanikidza.

Kune zvigadzirwa uye zvikwata zvekuchengetedza, chirevo chinodiwa chekuvaka: chengetedza rekodhi inotarisika yekuti ndiani kana chii chakapihwa simba, shanduko dzipi dzakaitika, marongero akashandiswa, uye kana nhume yakachinja chiyero. Izvi zvinogona kusanganisira zvakasainwa kana kuti append-chete provenance, kugona kudzikisira, kutariswa kwemuganho kwakajeka, uye kubvumidzwa kwevanhu kwezviito zvine chekuita zvakanyanya. Iro bepa haritauri kuita kumwe chete, asi rinopa chikonzero chechokwadi chekuyedza kuti gadhi wega wega anobvunza chiremera chekwakabva kwete chete nyika yazvino yakatorwa.

Chigumisiro chekudzorera chakakosha zvakaenzana. Sisitimu inogona kudzivirira kuuraya kunokuvadza uye ichiri kuita kuti kufambiswa kwebasa kusashande kana ikasapa nhanho yakachengeteka inotevera. Mumamiriro ezvinhu akashumwa ekuziva, zvikamu zvakavharika zvakazopera zvakachengeteka, zvichiratidza kuti kuteedzera uye kubatsira kunofanirwa kuyerwa pamwechete. Ongororo dzinofanirwa kurekodha zvikumbiro zvakarambidzwa, zviito zvakavharwa, kudzoreredza kwakachengeteka, basa risina kukwana, uye kukwesha kunoonekwa nemushandisi pane kudonhedza zvese kuita nhamba yekuramba kana yekutyora.

Kune zvakare chidzidzo chekuyera chezvichemo zveveruzhinji nezve kuchengetedzwa kweagent. Vanyori vakachengeta modhi, nharaunda yebasa, uye nzvimbo yekuita yakarongedzwa zvakakwana kuenzanisa mamiriro ekutonga zvakananga. Izvo zvinoita kuti iyo meshini igoneke, asi zvakare inoganhura izvo zvinogona kutaurwa nezve yakavhurika-yakapera ekugadzira masisitimu. Chiyero chebhenji kubva kune imwe dhizaini dhizaini haifanirwe kuratidzwa sechinhu chemuenzaniso wega, kunyanya kana mvumo, maturusi, kutema matanda, uye maitiro ekudzoreredza anogona kushandura mhedzisiro yakayerwa.

Interactive Mechanism

Interactive Mechanism: Iyo Inonyatsoshanda

Ongorora ari pasi tekinoroji kuseri kwekusimudzira uku uchipindirana.

Agent Lifecycle Stage:
1
User Intent & Planning: "Audit customer refund request #4092 and settle payment."
2
Tool Calling: Emits structured JSON call crm_get_transaction(id='4092').
3
Guardrail & Verification:🛡️ Paused: High-value action requires human operator sign-off.
4
Final Settlement: Refund recorded, email receipt dispatched, and audit log stored.
Core takeaway: An AI agent is not just a language model—it is a closed loop of planning, tool invocation, and environment feedback. Production systems require self-healing retries and strict human approval guardrails.
Interactive Concept Check+10 Points
AI Safety Quiz

What is 'specification gaming' in AI systems?

Zvekutarisa zvinotevera

Humbowo hunotevera hunofanirwa kuyedza kana maprovince-aware control anopona mamodheru akakura, mabasa asina kunyorwa, kuyedza kushandura kudzora nyika, uye zvipingaidzo zvekutumira. Iyo preprint inotsigira dhizaini hypothesis uye reproducible yekuongorora nzira; hazviratidze kuti chivakwa chegadhi chimwe chinogadzirisa kuchengetedzwa kwevazhinji-vamiririri.

Mapoka akazvimirira anofanirwa kumhanyisa iyo POLIS suite neyakaburitswa zvigadzirwa uye mishumo yemhedzisiro yemhuri yemuenzaniso, mamiriro ekutonga, uye chikamu chemhedzisiro. Kudzokororwa kunobatsira kwaizosiyana mataurirwo ebumbiro remitemo, marongero eshanduko, mutengo wekudzokera kumashure, uye ruzivo rwunowanikwa kune yekuteedzera layer. Vanofanirawo kushambadza kukundikana uye pedyo nekupotsa, kwete chete yekupedzisira yakaziva-kutyora mwero.

Deployers vanofanirwa kuyedza iyo imwechete musimboti vachipesana neyavo data inoyerera. Chiitiko chakachengeteka chinogona kuteedzera gwaro kubva kune yakatenderwa sosi kuburikidza nekupfupisa, kudhirowa, maturusi emafoni, uye kuyambuka-mumiriri nhume, wobva waona kuti zvirambidzo zvinoramba zvakabatanidzwa pazvinhu zvakatorwa. Muedzo unofanirwa kuvhara kubvisirwa, kupera nguva, zviremera zvinopokana, kuedzazve, kukundikana zvishoma, uye kuedza kunyora zvekare zvinyorwa pasina kushandura kwazvakabva. Marogi anofanirwa kuita kuti nzira yesarudzo inzwisise kune anoshanda.

Ongororo dzenguva yemberi dzinofanira kuongorora tradeoff pakati pekuvhara nekupedzisa. Murindi anomisa chikumbiro chese chisina kujeka anogona kuburitsa zero kutyora nekuramba basa rinobatsira, nepo gadhi anoregedzera achigona kuchengetedza kubuda nemutengo wekudonha chinyararire. Zviyero zvinobatsira kupedza basa, kuvharwa-chikumbiro chiyero, kukuvadza kwakaitika, nguva yekudzorera, mutoro wekuongorora kwevanhu, manyepo enhema, uye kana mumiriri anogona kusarudza nzira yakachengeteka yemukati kana nhume yaanofarira yarambwa.

Pakupedzisira, vaverengi vanofanira kuchengeta miganhu yepepa ichioneka. Icho chitsva, chisina-vezera-chakaongororwa preprint zvichibva pane zvakarongwa episodhi uye yakasarudzwa modhi yekupedzisira; haitauri kutyora kwepasirese, kuongororwa kwakazvimiririra, kana vimbiso yemapuratifomu emumiriri. Kutevera kwakasimba kwaizobatanidza kodhi yeruzhinji neyakanyoreswa kudzokorora, kugadzira-kufanana nemvumo magirafu, mitauro yakawanda uye multimodal mapindiro, uye bvunzo dzeanopikisa dzakagadzirirwa kunanga mafambiro pachayo.

Related guides & Quizzes

AI KuchengetedzaKuchengetedzwa kweAIAI AgentsLLM EvaluationsEdza zvaunoziva - edza yemahara AI quizTarisa kumusoro izwi reAI mune yedu glossaryTevedza iyo AI regulation tracker
Wakawana izvi zvinobatsira?