Pith Number

pith:RUSQMJ52

pith:2026:RUSQMJ52WHUMPZPUZT3YLACLWF

not attested not anchored not stored refs resolved

Prompting Policies for Multi-step Reasoning and Tool-Use in Black-box LLMs with Iterative Distillation of Experience

Ambarish Jash, Ketan Todi, Krishna Sayana

A reinforcement learning framework trains a lightweight prompter to optimize prompts for frozen black-box LLMs, lifting reasoning accuracy from 55% to 90%.

arxiv:2605.14443 v1 · 2026-05-14 · cs.AI · cs.LG · cs.MA

Open paper page JSON Open Graph Bundle Merged state Verified badge What is a Pith Number?

Add to your LaTeX paper

\usepackage{pith}
\pithnumber{RUSQMJ52WHUMPZPUZT3YLACLWF}

Prints a linked badge after your title and injects PDF metadata. Compiles on arXiv. Learn more · Embed verified badge

Record completeness

1 Bitcoin timestamp

2 Internet Archive

3 Author claim open · sign in to claim

4 Citations open

5 Replications open

✓ Portable graph bundle live · download bundle · merged state

The bundle contains the canonical record plus signed events. A mirror can host it anywhere and recompute the same current state with the deterministic merge algorithm.

Claims

C1strongest claim

We demonstrate significant gains, improving performance from 55% to 90% in logic-intensive reasoning and 74% to 91% in tool-use tasks. Furthermore, we analyze the structural evolution of prompts, demonstrating how the policy discovers specialized algorithmic heuristics.

C2weakest assumption

The lightweight prompter model can be optimized to maximize task-specific rewards for the larger frozen worker LLM using a contrastive experience buffer that couples scalar rewards with dense textual critiques.

C3one line summary

Iterative distillation of experience trains prompting policies that boost black-box LLM performance on reasoning and tool-use tasks from 55-74% to 90-91%.

References

58 extracted · 58 resolved · 1 Pith anchors

[1] Large Language Models as Optimizers 2023 · arXiv:2309.03409

[2] Deconstruct the golden response to understand the implicit steps, logic, and knowledge it used

[3] Prioritize Logic and Structure:For analytical, reasoning, or multi-step tasks, your improve- ments should focus on formalizing a step-by-step thinking process

[4] The new prompt should be self-contained

[5] What is the agent supposed to do?

Receipt and verification

First computed	2026-05-17T23:39:06.992092Z
Builder	pith-number-builder-2026-05-17-v1
Signature	Pith Ed25519 (`pith-v1-2026-05`) · public key
Schema	pith-number/v1.0

Canonical hash

8d250627bab1e8c7e5f4ccf785804bb14a9abc586acf3c34956be808ad920ef0

Aliases

arxiv: 2605.14443 · arxiv_version: 2605.14443v1 · doi: 10.48550/arxiv.2605.14443 · pith_short_12: RUSQMJ52WHUM · pith_short_16: RUSQMJ52WHUMPZPU · pith_short_8: RUSQMJ52

Agent API

Resolver JSON Graph JSON Events JSON Schema Signing key

Verify this Pith Number yourself

curl -sH 'Accept: application/ld+json' https://pith.science/pith/RUSQMJ52WHUMPZPUZT3YLACLWF \
  | jq -c '.canonical_record' \
  | python3 -c "import sys,json,hashlib; b=json.dumps(json.loads(sys.stdin.read()), sort_keys=True, separators=(',',':'), ensure_ascii=False).encode(); print(hashlib.sha256(b).hexdigest())"
# expect: 8d250627bab1e8c7e5f4ccf785804bb14a9abc586acf3c34956be808ad920ef0

Canonical record JSON

{
  "metadata": {
    "abstract_canon_sha256": "2b88354b4a55a7b8145f1894d1d768d6e3109b52e37c11ce0d8bbf93b62ed17d",
    "cross_cats_sorted": [
      "cs.LG",
      "cs.MA"
    ],
    "license": "http://creativecommons.org/licenses/by/4.0/",
    "primary_cat": "cs.AI",
    "submitted_at": "2026-05-14T06:38:19Z",
    "title_canon_sha256": "e4a37b171e149cf34fa11d843f3cb70333590f947323067315c69bf304ab40f2"
  },
  "schema_version": "1.0",
  "source": {
    "id": "2605.14443",
    "kind": "arxiv",
    "version": 1
  }
}