Slides: Continual Learning for AI Agents: From Failures to Durable Improvements - Soheil Feizi, RELAI
Source Video
Continual Learning for AI Agents: From Failures to Durable Improvements - Soheil Feizi, RELAI
Relationship To World's Fair 2026
These slides are extracted from a public AI Engineer YouTube video connected to World's Fair 2026. Speaker-matched clips are supporting context unless later confirmed as exact session recordings; official livestream recordings are day-level/event-level source material.
Related Scheduled Sessions
- No individual scheduled session mapping has been assigned yet; treat this as an event livestream deck.
Extracted Slides
OCR text:
Soheil Feizi
Founder & Chief Scientist, RELAI
Associate Prof, CS @ University of Maryland
https://relat.ai
OCR text:
Humans learn from experience.
vCiil
feedback act
World
act > get eee > improve without forgetting
OCR text:
Continual Learning for an Al Agent
ra
| AGENT |
MODEL HARNESS MEMORY
cae LLM(s): weights prompts - skills - in-session state }
; model selection tools - code - persistant ;
} workflow knowledge
continuously improve
the agent from its experiences ides
without forgetting. ane
y
= : PVC a ye wal elke) J
OCR text:
benchmark + evaluator
Teele ane lie Vetch Evaluator
curated task aU rab eth. 4 scores Output
PASS / FAIL / REWARD
+ Feedback
i
OCR text:
a raw log isn't feedback
errata ers An LLM / code analyzes the log
: PM alse abe oed lamer 6] sie [oL9
thd aT= 1a T- WAN SM ATO) AA tT: \ > So .
OCR text:
Matha oes] ?
An inferred distribution that replays what happened + what success means.
t Ba
ki Mocked / real tools Synthetic user Evaluators
feedback
run candidate agents against it, then keep the fix only if it passes. f
OCR text:
Three layers to improve the agent
oO Model
QO ada sty
S) Memory
A good learning engine asks for the oe aT Bigem + cS
OCR text:
Updating the model weights
SFT
imitate correct trajectories; needs labeled examples of the right behavior
Supervised fine-tuning
RL post-training
sample, score against a reward or preference signal, reinforce what wins
DPO-GRPO: RLVR
LoRA
limits the setof parameters that can change; cheaper, safer updates
Low-Rank Adaptation
'
They need:
Hard to apply to a raw production log (unless we lift it into a replayable enve
OCR text:
Updating the harness
Rewrite the prompts, skills, and code around the model.
Trace-to-harness GEPA & prompt search
ROC Oley nary Ob ane eGo Ran ay eaemet eo) bee aa an ren groban Tecan en pio aero meni Cava etic a Gre CMO Pe Omm Lento ars
pclae wis Src Cleo Cols mC HE ys a! Gai Canora cues ane Pcie Gaye retro ami) pie Clee Eco MR aerate Sc
OCR text:
Updating memory
Write down facts and distill skills, so the agent doesn’t rediscover them.
Information memory
store a fact or correction; e.g., “always confirm the date before booking”
Skill distillation
compress a successful trajectory into a reusable how-to packet
[scmetresviewmed asa patoafharess!
: : - works directly on (log + feedback) but usually unverifie F
OCR text:
Verifiable Continual Learning (VCL)
ee a eh ; .. improving an agent from its own experience, where every
han to help and to break nothing that already worked.
An executable test A measured delta A regression check
eee ae Serko s tre acer asic ro 3/4 ee ey rom a Om mCe he ca OMe mn teeSsea en Sen TEES Sea
acme eC Us Pie ees eC salm Biot ae aes rc) no
i
OCR text:
Replayable
Turn a one-off failure into a test you can re-run.
& learning environment
a ee ee a ce
symone persona, ‘epiuys the intwrachon
ae Pea Cine e
scorns pass /fa./ res
i
OCR text:
Holistic
One failure may have several causes and several possible repairs.
the agent cites a stale policy and skips the required escalation.
Route the fix to the layers that explains the failure withthe .. ... >
OCR text:
Lifelong
Anew fix must improve the new case without breaking the past.
© already optimized overE, .. E,. Anew failure E,.1 arrives.
a) Patch & hope > drift oO) Regression-aware learning
; : : : ' performance on E,s,
; ; oe no regression on £, .. &
Y .
Regression-control should be , not post-hor
OCR text:
Efficient
Efficiency in updates to the agent:
Model update
Search over harness , ,
Erna cn ee
Memory write , oe
A
Efficiency in regression-aware optimization loop
OCR text:
RELAI’s learning loop
oO Signals (logs - feedback - prompts)
© Replayable learning environments
:] Root-cause > route to a layer
oO) Regression-aware optimization
ry) Reviewable, versioned update : ,
OCR text:
Add VCL to your agents in
acRe- Seen anes
ot ] Use your own LLM
an Compatible with all major agent frameworks
Initial agent < : p
$ relai learning-env create --log-file --feedback
Create learning environments from lag/feedback or synthetically
; ; Simulators (mock/real tools, persona,...) and evaluators (code/LLMs)
y $ relai optimize
ae Holistic: adjusts prompts, models, tools, skills, s
Lifelong: online regression control i
Optimized version PR
c
OCR text:
Meridian Support Agent
Areproducible test-bed for continual learning in a tool-using support agent.
=) Asingle source of truth A) Interacting policies
a a ae aa AR acer a a et Caria ae eg an a A a a aera oc maaan iar
a) Deterministic evaluators Pe ie ee et
@ Decisions are tool calls Se a ee ee oe oe ee ee
co ) Regression-sensitive by design ST ae a Se
OCR text:
o
:a rude, adversarial caller
relai Tearning-eny create -- prompt
“A rade, adversarial culti tues casteter caewersution. phe caustentee devapes a”
dmautneri2ec Pigh dollar refune tngt snculd ret se grantee”
y
OCR text:
The generated
Persona,
intent, mocked/real
tools
ors kYAre ILM oazcl Orsi ce lac)
with feedback
all produced from one
interactive command.
4
OCR text:
Run it:the currentagentstruggles
relai simulate rude-user-multiturn-refund-escalation
0.78/1.00 average-two evaluators fail
WHEREITBREAKS
required-escalation
didnotroutetheunauthorizedrefundtoreview
latency-budget
toomanyturns/toolcallsunderpressure
WHATALREADYHOLDS
forbidden-direct-refund
held:neverissued therefund directly
safety-disclosure
held:no policy or contract leakage
OCR text:
The other source:
relal Joarning-env create \
Ome oun
“een fast elipiole refunds, nut do cot penerelize pererosity beyond
eee aie chee eee omen e | RM ra eee Ueno Occ
oO dat marsha 1c: (ol - ME ORa Tem EO Te CA EST SEO RD © a replayable learning env:
© feedback — th. cotrectinn eas Serre tas tac or tar are TESTOR SEIS | ag ora
OCR text:
Lifelong agent improvements
F
This is in practice:
each update is tested, every gain is measured, and nothing that already vw’ aa% *,
OCR text:
Takeaways
Agent continual learning is not only model fine-tuning.
ees eee ely ce Ceres orca tea) C Mee arehe rome OMe cio torre Be mane Gack trir sa Bren
Production logs are not learning environments.
ice hare t more me cs alm OMe CM mac pi WTO) CANO RCL an Marea COT
The frontier is regression-aware continual improvement.
Pra esate cf mean: VRC ae Gea) mone Chane om ronr ume nO 10 aam
, ,
Verifiable continual learning = + a as
Bi s'alaeleh
Slide-Derived Subjects To Review
Subject extraction uses video title, related session titles/descriptions, transcript context, and OCR text when available. OCR is best-effort and should be reviewed against the embedded slide images.