C.W.K.
Stream
Lesson 06 of 06 · published

Open Challenges in LLM Evaluation

~18 min · safety, research, frontier

Level 0Guesser
0 XP0/55 lessons0/10 achievements
0/150 XP to next level150 XP to go0% complete

The frontier problems

LLM evaluation is a young, fast-moving field. The hardest problems are still open. Knowing them keeps you honest about the limits of what you can promise.

Eight unsolved evaluation problems

  1. Evaluating creativity — how do you score a poem, a novel idea, a joke? Multiple valid outputs, taste is subjective, no reference.
  2. Evaluating reasoning — correct answer ≠ correct reasoning process. A model can pattern-match its way to the right answer without understanding.
  3. Evaluating safety at scale — adversarial attacks evolve faster than defenses. A safety eval today is a stale snapshot tomorrow.
  4. Long-context evaluation — how do you test 1M-token inputs systematically? Most benchmarks are short. Behavior at long context is qualitatively different.
  5. Multi-modal evaluation — scoring outputs that mix text, images, video, audio. Each modality needs its own framework.
  6. Agent evaluation — open-ended tasks with no single correct trajectory. How do you score "did this 47-step session do the right thing"?
  7. Meta-evaluation — making sure your metrics actually predict quality. The hardest meta-question.
  8. Benchmark contamination — public benchmarks all leak into training data eventually. The half-life of a benchmark's usefulness is shrinking.
Principle: You will not solve all of these. But understanding that evaluation is hard — and doing it anyway — puts you ahead of 90% of teams that ship LLM features. If you cannot measure it, you cannot improve it. Start measuring.

Where the field is going

  • More capability-eliciting evals — designed to surface latent abilities (and dangers) rather than measure performance.
  • More process evaluations — scoring not just final outputs but the reasoning trace.
  • More continuous evals — production traffic feeds back into the suite in real time.
  • More secret holdouts — held-back datasets that never appear in any training corpus.
  • More collaborative grading — humans + LLM judges + classifiers ensembled.

The eval mindset

The discipline is not "do I have a perfect metric?" It is "do I have an honest measurement, and am I improving it as the system improves?" Evaluation is a practice, not an artifact. Teams that internalize this ship faster, regress less, and trust their own systems.

Code

What a quarterly eval review meeting covers·markdown
# Quarterly eval review — agenda

## 1. Suite health
- Pass-rate trend across releases
- Per-tag breakdown — any tag bleeding?
- Cost per eval run (trending up?)

## 2. Suite drift
- Did we add new cases? How many?
- Are old cases still relevant or should we retire them?
- Production-distribution drift since last quarter?

## 3. Meta-eval
- Correlation of automated metrics with human ratings (sample 50)
- Correlation with user-effort signals
- Are any metrics now theatre?

## 4. Red-team status
- Last red-team run: when, who, findings
- Next red-team run: scheduled?

## 5. Tooling debt
- Anything broken / fragile in the eval pipeline
- Cost optimizations to land

## 6. People
- Is the team contributing cases? If not, why not?
- Onboarding for new joiners — who runs it?

External links

Exercise

Pick the open challenge from the list above that most affects your product right now. Write a one-page note on how your team's current evals fall short on that axis, and what you would build to close the gap. Share it with the team.

Progress

Progress is local-only — sign in to sync across devices.
Spotted a bug or have feedback on this page?Report an Issue

Comments 0

🔔 Reply notifications (sign in)
Sign inPlease sign in to comment.

No comments yet — be the first.