Skip to content
Bud's Wiki

WHALE — joint harness + weights optimization via alternating block ascent

TL;DR

Section titled “TL;DR”

WHALE is a follow-up to Meta-Harness that trains the LLM and the surrounding harness end-to-end by alternating block ascent: freeze the harness and post-train the model against it, then freeze the model and mutate the harness under a gradient-free proposer (an LLM emits candidate harnesses that a scorer ranks). The author’s central claim is that harness-only or weights-only optimization plateaus, and that switching legs when one bottlenecks creates new headroom for the other — WHALE beats harness-only, weights-only, and prompt+weights baselines “by a pretty wide margin” in the reported experiments. An adaptive variant monitors the training metric and switches phases at plateaus, matching a wide hyperparameter search over phase length.

Key claims

Section titled “Key claims”
  • WHALE alternates two update phases: (i) update model weights with the harness fixed, (ii) update the harness with the model fixed via a gradient-free LLM-proposer loop; both components can be the bottleneck at different points and switching legs creates headroom for the other [§thread, tweet 3].
  • WHALE outperforms three baselines — harness-only optimization, weights-only optimization, and prompt+weights joint optimization — by a “pretty wide margin” on the authors’ evaluation setup [§thread, tweet 4 / result figure].
  • An adaptive variant of WHALE monitors the training metric and switches phases when learning plateaus, and was competitive with a wide manual sweep over phase-length hyperparameters — removing the main scheduling hyperparameter [§thread, tweet 5].
  • Author’s forward bet: large post-training runs will benefit from spending ~1% of the rollout budget on harness improvement, framing “text optimization” (harness search) as a first-class post-training axis rather than a scaffolding detail [§thread, tweet 6].
  • Named open issue: WHALE “co-overfits” to one good harness rather than producing a model robust across harnesses; the author calls this out as the next problem alongside training models to explicitly write/improve their own harness with access to internal activations [§thread, tweet 7].

Method

Section titled “Method”

The block-ascent recipe as described in the thread: pick an initial harness, post-train the LLM against it (leg 1 — standard weights update), then freeze the LLM and run a gradient-free harness update (leg 2 — an LLM generates candidate harnesses that are scored by rollouts and the best is selected, i.e. the “Meta-Harness”-style proposer loop). Iterate. The adaptive scheduler switches legs at plateaus rather than at fixed intervals. The thread describes the acronym as WHALE without expanding it; the earlier Meta-Harness thread is cited as the setup this generalizes.

Results

Section titled “Results”

Reported qualitatively only in the thread (no absolute benchmark numbers): WHALE beats (a) harness-only, (b) weights-only, and (c) prompt+weights baselines by a wide margin on the reported task family, and the adaptive-scheduler variant matches a manual phase-length sweep. Attached figures show a bar chart of WHALE vs the three baselines and a training-curve figure demonstrating the plateau-triggered leg switches; specific numeric deltas and the evaluation benchmark are not stated in-thread.

Why it’s interesting

Section titled “Why it’s interesting”

WHALE is a concrete answer to the wiki’s long-standing open question — restated most sharply on AI-for-AI Research and in AutoDesign: Meta-Harness Optimization for Long-Horizon Agentic Design‘s meta-harness thesis — of whether “next-round improvement from re-injecting prior-round outputs” produces real headroom or just co-adapts the harness to a fixed model. WHALE claims yes, but only if you alternate: freezing one side while updating the other changes what the other side can optimize, so the two-leg schedule creates headroom neither leg has on its own. This complements Ornith-1.5: From Self-Scaffolding to Self-Improvement, which jointly optimizes task-proposer + harness + policy under a single GRPO objective at training time; WHALE isolates the weights ↔ harness pair specifically, uses gradient-free updates on the harness side, and reports the alternation itself as load-bearing rather than the single-objective coupling. It also sharpens the “harness carries transferable capability” finding in AutoDesign: Meta-Harness Optimization for Long-Horizon Agentic Design (DesignHarness transfers across 7 model/agent pairs for +12.4 pts) with a co-training variant of the same insight — with the caveat, flagged by the author, that WHALE-style co-training currently co-overfits the model to one specific harness, which is the exact opposite of AutoDesign’s cross-pair transferability. The concept page’s TL;DR calls the harness-transferability question open; WHALE argues alternation raises the ceiling but at a robustness cost.

See also

Section titled “See also”