FA-98491 / Beam search bookkeeping / Open access

Beam search: length normalization changes accumulated search scores · case 01

length normalization changes accumulated search scores.

Verified by executionVariant 1 · 6 checks per implementationDownload source bundle ↓JSON ↗

ROOT CAUSE

The decoder feeds a ranking-only adjustment back into subsequent token accumulation.

THE FAILURE

The decoder feeds a ranking-only adjustment back into subsequent token accumulation.

Unsuccessful approach: Keeping only normalized values and adding new token deltas still corrupts later ranks.

Case contract

Given token deltas and alpha, return prefix raw totals; alpha is reserved for final ranking.

Why this case matters

Beam decoders rely on consistent hypothesis state while expanding and ranking candidates.

1 / The failure

Exit 1

"""Failure Map reference implementation. Python standard library only."""
import json

N = 1
observations = []
def solve(deltas, alpha):
    return [sum(deltas[:i+1])/(1+alpha*(i+1)) for i in range(len(deltas))]
def check(label, actual, expected):
    observations.append({"check": label, "actual": actual, "expected": expected, "passed": actual == expected})
check('raw prefix totals',solve([-N,-2],1),[-N,-N-2])
check('single step',solve([-3],N),[-3])
check('empty steps',solve([],N),[])
check('positive toy deltas',solve([N,1],2),[N,N+1])
check('later negative step',solve([0,-N,-1],.5),[0,-N,-N-1])
check('zero penalty',solve([-1,-1],0),[-1,-2])
print(json.dumps({"observations": observations, "passed": all(x["passed"] for x in observations)}, ensure_ascii=False))
raise SystemExit(0 if all(x["passed"] for x in observations) else 1)

Boundary fixture	Actual	Expected	Outcome
raw prefix totals	`[-0.5, -1.0]`	`[-1, -3]`	Failed
single step	`[-1.5]`	`[-3]`	Failed
empty steps	`[]`	`[]`	Passed
positive toy deltas	`[0.3333333333333333, 0.4]`	`[1, 2]`	Failed
later negative step	`[0.0, -0.5, -0.8]`	`[0, -1, -2]`	Failed
zero penalty	`[-1.0, -2.0]`	`[-1, -2]`	Passed

SHA-256 / d2090e4dc4c24d119d3b46f0a9309cd8be69982064d14ca1c8fe63b2a1bfb28f

2 / The unsuccessful fix

Exit 1

"""Failure Map reference implementation. Python standard library only."""
import json

N = 1
observations = []
def solve(deltas, alpha):
    return [sum(deltas[:i+1])/(1+alpha*i) for i in range(len(deltas))]
def check(label, actual, expected):
    observations.append({"check": label, "actual": actual, "expected": expected, "passed": actual == expected})
check('raw prefix totals',solve([-N,-2],1),[-N,-N-2])
check('single step',solve([-3],N),[-3])
check('empty steps',solve([],N),[])
check('positive toy deltas',solve([N,1],2),[N,N+1])
check('later negative step',solve([0,-N,-1],.5),[0,-N,-N-1])
check('zero penalty',solve([-1,-1],0),[-1,-2])
print(json.dumps({"observations": observations, "passed": all(x["passed"] for x in observations)}, ensure_ascii=False))
raise SystemExit(0 if all(x["passed"] for x in observations) else 1)

Boundary fixture	Actual	Expected	Outcome
raw prefix totals	`[-1.0, -1.5]`	`[-1, -3]`	Failed
single step	`[-3.0]`	`[-3]`	Passed
empty steps	`[]`	`[]`	Passed
positive toy deltas	`[1.0, 0.6666666666666666]`	`[1, 2]`	Failed
later negative step	`[0.0, -0.6666666666666666, -1.0]`	`[0, -1, -2]`	Failed
zero penalty	`[-1.0, -2.0]`	`[-1, -2]`	Passed

SHA-256 / 667e568a11f4d867a79673259abcc77ac3aa8e64d431b60c237bbdc5502d183d

HELD IN THE MEMBER ARCHIVE

The verified repair and its recorded checks are member-only.

This mechanism has 6 recorded checks per implementation. The open-access tier publishes the failure and the unsuccessful fix; the repaired source that passes every check, and the observations that prove it, are available to members.

Every case sharing this mechanism uses the same contract and the same repair, so this one record is held back for all of them.

Member access is invitation-based. Sign in with your invited account to inspect the repair.

Verification & scope

This is a finite toy beam-state contract and makes no claim about neural model quality or production decoding. This reproducer isolates one failure mechanism. Results cover the supplied fixtures. Variants within a family share a test contract and should remain grouped when constructing evaluation splits. Related mechanisms with a shared evaluation_group must also remain together; these controlled models are not independent production incidents.

Observations recorded using Python 3.12.14 at 2026-09-30T16:48:13.668730+00:00.

Case digest / 054ba06058dc7a070eca5523c5e882ec9aeca6b2c23422fdf269fb557d57dd7b